NVIDIA GPUs have become so in-demand for so-called "AI" workloads that a "black market" has emerged around them -- at least, in the eyes of the US Government. In China, it's simply a "market." We adventured on extensive travels throughout Asia and spent hundreds of hours investigating the issue

The Highlights

• We spoke to everyone about NVIDIA’s AI GPU black market, including middlemen who connect buyers and suppliers
• The export of these GPUs to China is in violation of US Government law, which includes numerous restrictions on semiconductor processing capabilities
• Fueling greed, manipulation, and propaganda, we think NVIDIA is playing all sides

Table of Contents

• AutoTOC

Intro

We’re multiple administrations deep in a technological cold war over processing power between the United States and China. China’s Cyberspace Administration has labeled some US graphics processing products as a security risk, seeking answers about US government backdoors in the silicon. Meanwhile, the United States has imposed heavy restrictions on exports of graphics processing units, or GPUs, being sold to Chinese companies by American companies. The sale requires rarely-granted licenses for each import scenario to legally export GPUs above a certain performance level, with the stated objective being to restrict progress of private and government projects, including AI development, in China, while trying to maintain the US’ claimed AI leadership. 

The United States takes this so seriously that, just this week, the Department of Justice had two Chinese nationals arrested in California for what the DOJ alleges is the smuggling of tens of millions of dollars of GPUs.

But where there’s prohibition, there’s smuggling.

Editor's note: This was originally published on August 17, 2025 as a video. This content has been adapted to written format for this article and is unchanged from the original publication. This particular story had a saga of what we felt was censorship behind it, thanks to Bloomberg L.P., which we've detailed extensively: Part 1 (Our Channel Could be Deleted) and Part 2 (YouTube's Systematic Punishment). This was written to be seen as a video, so the adaptation sticks to a more viewable/colloquial use of language.

Credits

Host, Writing, Lead Editing

Steve Burke

Editing

Vitalii Makhnovets
Tim Phetdara

Editing, Graphics

Andrew Coleman

Camera

Tannen Williams

Research and Writing

Ben Benson

Acquiring "Illegal" GPUs in China

In China, it's not an illegal market -- it's just a market.

We spoke to everyone about this NVIDIA AI GPU black market: We found middlemen who connect buyers and suppliers, users who can understand the demand and explain it, using the most dystopian definition of wealth -- how many GPUs one has.

We also found independent repair shops who, simply doing their jobs, salvage valuable silicon components from dead PCBs of banned GPUs, innovatively hand-modifying them to be better than and have more VRAM than NVIDIA’s own official product SKUs. These shops are not explicitly a part of any "black market," they're just repair shops that happen to sometimes work on export-controlled GPUs.

We met multiple people who, when asked the same question, gave the same passcode-like Chinese idiomatic expression, or chengyu (成语), which translates to “open one eye, close one eye.” In other words, it means, “to turn a blind eye.” We heard this saying so much that we made a T-shirt based on it to help fund this investigative report.

"There's a new kind of black market, and it's high-end AI GPUs. This particular black market is worth billions of dollars a year"

And among others, we even spoke with a US-based Chinese national buying video cards to strip them and ship the GPUs to Chinese companies, which violates US export control law.

One of our viewers was able to meet with a GPU smuggler, whom we’ll call “The Plug.” The smuggler spoke limited English, but they both understood one universal truth: money. He operated a GPU testing rig inside of a Prius that he drives around multiple states in the Western United States. The least suspicious thing in his car was a spare license plate in the trunk, but we’ll come back to “The Plug” towards the end.

"Black market" is normally a phrase associated with drugs or guns, but there's a new kind of black market: high-end, AI GPUs. This particular black market is worth billions of dollars a year and is hiding in plain sight.

Hong Kong is our first stop along our journey. 

By skyscraper count, Hong Kong would be the tallest city on earth. The density is unbelievable. We spent a few days here for this story, wandering markets and meeting sources.

We went to Hong Kong to learn what the demand drivers are for the banned GPUs, how they get into China, and more about the illicit side of the GPU smuggling business. 

Big Adventure 

To provide a big overview of our big adventure, it began when we booked a 24-hour plane ticket to Hong Kong and 20 days of hotels across Hong Kong, Shenzhen, Zhengzhou, Huizhou, Taipei, and more. At least one of us got detained by at least one of the governments involved in this story, but we can't talk about these indecipherably singular or plural instances or instance of any aforementioned detainment.

Then we talked to a lawyer about the previous sentence, then investigated whether or not GPUs are actually smuggled with lobsters into Hong Kong.

We eventually found a pile of GPUs that are now export controlled and met a professor who really wants to make sure you know that they were “legally obtained.”

After leaving Hong Kong, we booked a boat to Shekou, China. Then headed to Shenzhen, got kicked out of a warehouse, and used high-speed rail to go deep into China, where we met a guy who thinks that desoldering a GPU and reballing it is no big deal. 

We got a ton of information about GPU smuggling from a guy named “5.” 

After that, we went back to Shenzhen for the third time in two weeks, went back to Hong Kong, flew to Taiwan, before finally getting back on the plane to the US.

Perspective

We shot over 12 hours of interviews

We spoke to a lot of people on the record for this story and learned about the dystopian world of high-tech GPU smuggling. We spoke to people ranging from owner-operator trading companies to professors of economics building datacenters. We even tried to talk to the US Department of Commerce, but they didn’t reply, and every person on the chain from the US Department of State had out of office auto responders because we coincidentally emailed them the same week of an especially problematic story that enigmatically involved namedropping. 

We wanted to connect as many pieces of this puzzle together as we can today, and that’s why we shot over 12 hours of interviews that we just spent weeks cutting down. The point is finding people who know people, and each person in our lineup today led us to at least one other person in this video, eventually building the full pipeline of smuggler-to-user. Although this story isn’t about drugs, it is about a different thing that billionaire executives get a high from: AI. 

But before we get to any of those interviews, we need to establish the basics of this geopolitical mess.

The story is complicated, so we'll start with defining these key facts: Why these GPUs are banned, the new 15% license (which only applies to two GPU models), who buying and selling is legal or illegal for, and then the timeline. 

We have over 400 pages of research that went into finding sources and understanding the laws.

Why They’re Banned

AI GPUs have been in the mainstream news constantly.

This story has been a complete mess to follow. It has spanned years and two US administrations. We have over 400 pages of research that went into finding sources and understanding the laws. 

Reports warned of NVIDIA product use in nuclear weapons research

Here’s why governments care about AI GPUs:

NVIDIA functionally holds a GPU monopoly in our sector of the industry, which is building computers to play video games. That now feels relatively innocent by comparison to AI. The company leveraged decades of gaming domination to build a foundation for what it now focuses on, which is making the most powerful GPUs for AI in the world. Reports warned of NVIDIA product use in nuclear weapons research, facial recognition technology allegedly used in Russia to suppress dissent and growing concerns of AI facial recognition use in the US for similar deployments, alongside reports of use in international spying and in drone warfare. 

NVIDIA finds itself in the middle of all of this. Even though NVIDIA disputes selling to some of these entities -- for example, it says that it doesn’t sell GPUs to Russia -- the products still find their way there. NVIDIA is making money one way or the other. Someone is buying it, maybe from someone else who bought it from someone else, and it may be transacted through smuggling. But NVIDIA does end up selling the device ultimately to somebody. Regardless of who they sell to, NVIDIA plays a big part in this worldwide obsession of AI.

And we think it’s playing all sides, but we’ll talk about that more at the end. Besides, when there’s a gold rush, it’s better to sell the pickaxe than swing it.

Although the US doesn’t talk too much about its own use of AI, it spends a lot of time talking about China’s.

The US restricts NVIDIA’s GPUs through export control rules that ban the sale of certain GPUs into China. The restriction is for the sale of GPUs by American companies or companies that want to transact business in America to companies that are in China or the Chinese government itself. Some examples of export-controlled GPUs include gaming GPUs like the RTX 5090 and 4090, which are useful in AI applications (mostly for their high VRAM capacity), and data center/AI GPUs like the A100, H100, H200, and B100, as well as the others shown in the image above. This list is constantly in flux. There are some new and incoming exceptions for the NVIDIA H20 specifically, which has faced Schrodinger’s GPU ban depending on whether CEO Jensen Huang had a one-million-dollar dinner at Mar-a-Lago with Donald Trump on a given week. 

GPU Export Control Timeline

Now, we’ll get into the timeline of GPU export controls across multiple administrations. We have a separate article with the full, bulleted timeline here.

Obama Administration and Early AI Talk

Even at the end of the Obama Administration in 2016, the US Government was just starting to talk about AI in relation to national security in a Wired interview. Then-President Barack Obama said:

“Developing international norms, rules, protocols, verification mechanisms around cyber security generally and AI, in particular, is in its infancy. You got a lot of non-state actors who are the biggest players. Part of the problem is that identifying who's doing what is much more difficult. If you're building a bunch of ICBMs, we see them. If somebody's sitting at a keyboard, we don't. And so, we've begun this conversation. A lot of the conversation right now is not at the level of dealing with real sophisticated AI but has more to do with essentially states establishing norms about how they use their cyber capability. Who are you more afraid of: big brother and the state or the guy who's trying to empty out your bank account? Part of the reason that's so difficult is that if we're going to police this wild west, whether it's the internet, or AI, or any of these other areas, then by definition, the government's got to have capabilities. If it's got capabilities, then they're subject to abuse. And, at a time when there's been a lot of mistrust built up about government, that makes it difficult.”

First Trump Administration and Biden Administration

Those were the early days.

the Biden Administration took major action in 2022 by restricting exports to China, Hong Kong, and Macau

The first Trump Administration started a commission advising Congress on maintaining AI leadership, including simply banning the sale of some advanced semiconductor equipment and chips to China. Years of back-and-forth, a pandemic, an election, and Chat GPT’s launch later, then the Biden Administration took major action in 2022 by restricting exports to China, Hong Kong, and Macau. 

NVIDIA’s Ampere architecture A100 and newer codename “Hopper” H100 GPUs and systems were all restricted. NVIDIA shed tears for $400 million worth of lost sales as a result and was especially sad when many of its export-compliant alternatives to these also got banned, like its newly-created A800, H800, and L40S, in addition to NVIDIA’s RTX 4090 gaming card. NVIDIA said it didn’t expect “near-term meaningful impact” on its financials.

Second Trump Administration

NVIDIA responded by designing a workaround to the previously worked-around workaround, leading to the H20. Then the government added new rules for high memory bandwidth cards and the Biden admin tried to come up with an AI chip diffusion rule that would limit the quantity of GPUs being sold into different countries rather than only by processing power metrics, because the government really didn’t know how the fuck to measure these things and NVIDIA, knowing more about GPUs, could tweak any dial it wanted to just barely be compliant. Then DeepSeek came out and everyone panicked, stocks plummeted, and the government scrutinized the role of NVIDIA GPUs in it. In February 2025, Fiscal Year 2025 results were posted and NVIDIA’s Singapore revenue skyrocketed to 18% of total revenue based on customer billing location despite shipments to Singapore being claimed to be less than 2% of Fiscal Year 2025 revenue, which caused people to say “wait a minute.” Unrelated: Several GPU smugglers were arrested in Singapore one day after the fiscal year report was posted, which caused people to say “that makes more sense.”

In May, 2025, Trump implemented wide-sweeping tariffs and rescinded the Biden chip diffusion rule that would have limited how many AI GPUs Jensen could sell to other countries

AMD spawned out of nowhere to say it wrote-down $800MM of inventory due to export controls. NVIDIA one-upped it with a write-off of $5.5 billion.

In May, 2025, Trump implemented wide-sweeping tariffs and rescinded the Biden chip diffusion rule that would have limited how many AI GPUs Jensen could sell to other countries, Jensen then said, “It’s just an incredible vision. I think this is going to be a transformative idea for the next century for us. These 2 initiatives are completely visionary and it’s going to be transformative for America.”

The H20 GPU that was created to comply with rules was still okay, then Jensen had dinner with Trump at Mar-a-Lago for $1,000,000, then the H20 got banned -- he must have chosen a bad restaurant. In July, Jensen Huang met with Trump and was permitted to sell H20s again, Huang went to China, China said Jensen’s GPUs have tracking devices and backdoors, NVIDIA denied that, Jensen went to Washington, Trump then spoke very highly of Huang (and Lisa Su), and more importantly, he ragged on Intel’s CEO. Intel’s CEO then went to Washington, so Trump likes him now. Tim Cook materialized from the infernal plane to give Trump a 24-karat gold “gift,” and “gift” is in quotes because that’s not what that’s called. Then we get to this past week, when Trump asked NVIDIA and AMD to pay 20% to the US government for sale SPECIFICALLY of the H20 and AMD Instinct MI308 sales, not all GPUs as some erroneously reported. Cousins Jensen Huang and Lisa Su negotiated Trump down to 15%, and now they’re allowed to sell two specific cards that were originally created to comply with the laws before they changed and somehow everyone walks away a winner. Except now China doesn’t want them anyway.

That about sums it up.

Except one last thing that happened as we were writing this: The Department of Commerce doesn’t yet know the legality of the deal, with Tom’s Hardware highlighting legal expert arguments over Article I, Section 9 of the Constitution.

If you want the full details with all of the in-between, make sure to check out our massive timeline article. 

A New Law

There was also recent news about a 15% revenue share between NVIDIA and AMD with the United States Government for sale of some AI GPUs. Trump stated, “This is an old chip that China already has and I deal with Jensen who is a great guy and NVIDIA. The chip that we’re talking about, the H20, is an old chip. China already has it in a different form, different name, but they have it. Or they have a combination of 2 will make up for it and even then some […] but the H20 is obsolete. You know, it’s one of those things, but it still has a market. So I said, listen, I want 20% if I'm going to prove this for you, for the country, for our country, for the US. I don't want it myself, you know, every time I say like 747, I want. Yeah, for the Air Force. So when I say I want 20, I want for the country. I only care about the country. I don't care about myself and he said ‘Would you make it 15?’ So we negotiated a little deal. So he's selling an essentially old chip that Huawei has a similar chip, a chip that does the same thing and I said 'good, if I'm going to give it to you' because they have a, you know, they have a stopper, what we call a stopper. Not allowed to do it. A restricted is really known as a restrictive covenant.”

If you were to read only the headlines, you’d think that this applies to all GPUs and that the ban is over, and then you might also think that a market in China, or in the US’ eyes, an illegal market in China, would cease to be so illegal. That’s not the case. 

This new 15% revenue share would, if it’s legal (and they’re not sure yet), allow NVIDIA to sell specifically the NVIDIA H20 GPU to approved Chinese entities. They likely can’t be on the entity list. It would allow AMD to sell specifically the Instinct MI308 GPU to approved Chinese entities. The proposed license would not affect any other GPU that currently does not have a license. The government hasn’t made clear yet if NVIDIA’s partners would also be permitted to make these sales. 

That means that other banned GPUs, including the RTX 4090, RTX 5090, H100, B100, B200, and so forth, remain banned. 

The H20’s extremely high 96GB memory capacity would enable large models to fit in memory and actually run, especially with multiple GPUs in a single rack, even if it’s slower. That means companies can achieve performance targets by stacking GPUs which are lower clock and core count but higher capacity.

As for the newer Blackwell architecture GPUs, Trump stated, “Now Jensen also has, Jensen’s a very brilliant guy, and Jensen also has a new chip, the Blackwell. Do you know what the Blackwell is? The Blackwell is super duper advanced.”

Let’s not give NVIDIA any ideas on new GPU names.

Trump added, “I wouldn’t make a deal with that. Although it’s possible I’d make a deal, a somewhat enhanced in a negative way Blackwell. In other words, take 30% to 50% off of it. But that’s the latest and the greatest in the world. Nobody has it. They won’t have it for 5 years. On the Blackwell, I think he’s coming to see me again about that. But that will be an unenhanced version of the big one.”

For now, these “Super Duper” GPUs are not licensed for sale. 

In short: The H20 and MI308 were compliant with the US Government’s original rules, then the rules changed during design & production and they were banned, then AMD and NVIDIA collectively declared over $6B in financial impact as a result. 

Then the US Government said, ‘Wait a minute, we can help with that if you cut us in.’

It reminds us of the unpredictability that we highlighted in our tariffs documentary.

The H20 most certainly is not “obsolete,” though. It’s still very desirable in China, and with a lot of them, they become particularly potent.  

How the US Determines GPU Bans

Setting a threshold for banned hardware should be objective since it can be tested.

This is a graph from the Department of Commerce that visualizes the original threshold at which a computing product became automatically banned for export to parts of the Middle-East and China without a granted license. There have been some changes since, but back when this was made, the “Total Processing Performance” score on the Y-axis was used to determine cards in need of a license. The government needed a metric to calculate against, so it created its own.

Accelerators and video cards have a lot of metrics in their spec sheets, including memory capacity (which is critical and as simple as a pass/fail for certain training and AI uses), memory bandwidth, GPU clock speed, GPU SM or CU count, TPCs, Tensor Cores, ROPs that are sometimes randomly missing on NVIDIA devices, TFLOPS, TOPS, PFLOPS, GFLOPS, gigabits, power, and more.

So then, banning a product could probably be based on some sort of benchmark rather than a random metric from a spec sheet, but the US government, illustrating what an absolute clusterfuck this situation was and now remains, decided to instead multiply one random metric from a specsheet against the bit length of the operation being executed. FLOPS, or Floating Point Operations Per Second, and TOPS, or Tera Operations Per Second, are calculated by the company making the spec sheet and aren’t a great measurement of actual performance. These numbers are based on both marketing and whether we’re talking about FP8, half-precision FP16, single-precision FP32, double-precision FP64, or Tensor performance, so the government said, “the rate of MacTOPS is to be calculated at its maximum value theoretically possible” and “the rate of MacTOPS is assumed to be the highest value the manufacturer claims in brochure[s] for the integrated circuit.” So it’s not based on a bunch of real-world benchmarks of applications or something useful. The government also references MacTOPS as the theoretical peak of TeraOPS in multiply-accumulate computations.

The Biden administration used this “Total Processing Performance” (or TPP) score in 2023, with the government later adding a “Performance Density” metric dividing the TPP by the die area in square millimeters. In other words, the government didn’t want NVIDIA to be able to sell more of a lower-performance GPU to make-up for the loss of high-performing parts with multi-GPU solutions. 

There are a lot of reasons this doesn’t capture the full picture, like sparsity, APIs, differing methods to calculate FLOPS, and different performance for different applications, but the government needed a way to define a threshold, so this is what it made. The limit was a TPP score of 4800, exceeded by even the RTX 4090 when calculated using Tensor performance.

Now, if this doesn’t mean anything to you, that’s OK, because it probably doesn’t mean anything to people signing the laws either. Or maybe that’s not OK, but you get the idea.  

Something like a higher memory capacity, lower FLOPS performance GPU or even series of GPUs like RTX 3060 12GB cards might be able to get the work done more effectively if it only needs memory. Memory wasn’t factored into the TPP calculation.

That was the point of the H20, but then the absurdity of the situation expanded by introducing an opaque memory bandwidth requirement. The Register wrote, “Unlike with previous export controls, the Commerce Department's Bureau of Industry and Security (BIS) hasn't issued specific guidance on how much I/O or memory bandwidth is too much.”

So it seems that, across now two administrations, the United States is creating formulas based upon numerical calculations and then, when NVIDIA and AMD tweak numbers to fit within that box, it is retconning those rules in a guess-and-check process.

Experts & Roles

We spoke to a lot of people in this story across different languages, which made it complicated. The discussions come from:

Dr. Vinci Chow, Senior Lecturer in the Department of Economics at the Chinese University of Hong Kong, responsible for building his department’s machine learning servers and sourcing GPUs from middlemen suppliers.

Dr. Zǐ Háo Fù, a Research Assistant Professor at the Chinese University of Hong Kong and Affiliated Lecturer at the University of Cambridge. Zihao specializes in both computer science and linguistics.

Various retail workers at the computer markets in Hong Kong to get a ground-level understanding.

Zhou, Creative Director of Product at video card manufacturer Yeston. Yeston is not involved in any "black market" of GPUs and mostly deals with AMD; however, the company gave us a tour of its GPU factories to better understand production processes.

An anonymous seller who goes by the pseudonym of “SILVER,” based in Shenzhen Bao’an and manages a warehouse in Hong Kong that receives and processes smuggled GPUs.

Vincent, a resourceful fence who lives next to the warehousing and markets filled with GPUs and accelerators. He buys and sells these devices in Shenzhen Huaqiangbei. His job is to know people. 

Vincent’s Cousin is the fixer’s fixer. If someone needs a component to fix a broken video card or to build a new one, they go to people like him to get the integrated circuits.

“Mr. 5,” a Bilibili hardware reviewer with a specialized focus on thermal solutions. Like us, Mr. 5 has had run-ins with NVIDIA that have ended in a soured relationship over disputes regarding independent reviews and editorial independence.

Brother Zhang, a renowned Bilibili uploader (basically a YouTuber in China) who runs a video card repair shop in Zhengzhou. In addition to repairs, he regularly gets large orders from customers asking him to build them custom, unofficial, higher VRAM capacity NVIDIA GPUs for large language model tasks. Brother Zhang is not a direct part of any "black market," as, again, his business operates legitimately within China as a repair shop. It does, however, come into contact with GPUs the USA considers to be "illegal" for sale into China.

Companies in Singapore and Taiwan who act as intermediaries between NVIDIA, NVIDIA’s partners, and companies in China. The Singaporean and Taiwanese companies are able to bring banned GPUs and servers in and re-sell them to Chinese companies, skirting export controls. We are unable to disclose their identities as there would likely be punishment from multiple governments and NVIDIA.

The “Plug,” a US-based Chinese citizen who drives around the country buying hardware from American end-users and resells it to companies in China and Hong Kong.

And a special thanks to our translator Raymen Wu of the BLK SODA agency in Taiwan.

The Smuggling Pipeline

Here's the pipeline.

If a GPU doesn’t fall off the back of a truck in China after it was assembled, or if it isn’t a "QC defect" that disappears from the scrap pile, it may instead be moved by “ants” to get to China. There’s a saying in Chinese that we heard a few times (mayi banjia - 蚂蚁搬家), which translates to “ants moving” that represents a linear pipeline. Each ant in the fireline serves a specific role. It’ll help to name those to keep everything straight. Here’s what we came up with:

The Source has access to GPUs. This could be as innocent as you unknowingly selling your card on Facebook Marketplace to The Plug. 

The Plug is responsible for acquiring from the original source and re-selling the hardware to the China-based distributors. 

In between, there’s a Mule. Sometimes this is the Plug himself doing a trip home. We learned that overseas students also regularly return with what are feasibly defended as personal GPUs, that they bought at a retailer like Best Buy, that may then get resold for markup and profit. In either case, the Mule gets the GPU into the country either by shipping it without interception or by hand carrying it.

Next is the Middleman, receiving the GPUs and often interfacing with or managing the warehouses that store the cards. The Middleman buys from multiple Plugs, including factories that get rid of rejects with fixable or unimportant QC defects, then sells those devices to more localized distributors.

That’s when we get to the Fence, who buys and sells GPUs between middlemen and warehouses to end users in China.

Then, we have who we’re calling the Fixer, except this time, it’s literal. The Fixer is an optional step that may involve soldering and modifying a GPU to improve it beyond its original specification to make it more marketable for domestic AI uses. They might also just fix QC defects from the factories.

Finally, we have the User. This is self-explanatory: The user is the demand driver, and oftentimes, large enterprises want dozens or hundreds of GPUs or more, while smaller users like the university may just want individual units or small batches.

With everyone’s role named, let’s continue.

Demand Drivers

Our journey began in Hong Kong. 

HONG KONG: Dr. Vinci Chow, Chinese University of Hong Kong

Upon arriving in Hong Kong, one of our GPU dealer informants slipped us a price sheet for a mix of GPUs that are both export controlled and not.

Some of the banned ones on the list include the 5090, 4090, 4090 (48GB), A6000, A100, and A100 (80GB). The person who provided the price sheet works with smugglers. The problem is, we don’t know what the going rate for these GPUs is.  

So we hopped in a Hong Kong cab and headed across Victoria Harbor and over to the Chinese University of Hong Kong to meet up with Dr. Vinci Chow.

We found him through a Reuters story from a couple years ago. He works in the economics department at the university and he’s responsible for having built many of the servers and machine learning systems that are in use daily. 

"Upon meeting Chow, he showed us some A100 GPUs, which he emphasized were, 'legally obtained.'"

He has a whole blog detailing his process, including the difficulty of sourcing components and the ease with fixing problems when so close to Shenzhen, such as having custom PCIe riser cables built and basically done the same day. 

Dr. Vinci Chow is the right person to start us off with this story and educate us on where to go next. We’re here to understand the user’s perspective because before there could be any market for it, there has to be demand. And that is what the university and organizations like it generate. 

it's illegal to export to China without [a] permit but it's completely legal on our side, right? There's nothing that says that you cannot buy a high-end GPU.

- Dr. Vinci Chow

Upon meeting Chow, he showed us some A100 GPUs, which he emphasized were “legally obtained.” He elaborated that he obtained them around when ChatGPT was released and got them for “quite cheap.” He expressed that, during this time, even the A100, which became export-banned in 2022, “was actually quite cheap.” He shared that he got the A100s for “less than 10,000 US Dollars.” Following ChatGPT’s release, however, Chow said that the prices went “crazy.”

Pointing out the oddity of the export ban, Chow stated, “It's interesting because, you know, it's the export ban, right? So it's illegal to export to China without [a] permit but it's completely legal on our side, right? There's nothing that says that you cannot buy a high-end GPU. So, from our end, as long as we follow all relevant procedures, there's absolutely nothing illegal about, for us over here, to buy these GPUs. So that makes it for a very interesting environment.” He added, “These universities have been purchasing GPUs. It’s completely legal on our side, right? Yet this is clearly not supposed to happen from the perspective of the US government.”

Regarding pricing and availability of export-banned GPUs, the professor stated, “If you pay enough, supply is there. Maybe not if you want to build a super cluster, right? For research, most researchers are talking about 1 or 2 GPUs. And if they have the funding, then it’s possible to obtain that. It’s just everything is more expensive.”

Having built Chinese University of Hong Kong’s cluster of GPUs during COVID, he has first-hand knowledge of how much banned GPUs cost to get in Hong Kong and estimated that an H200 would cost around “$30K.” When we looked at our GPU price sheet, the H200 was going for 213,000 HKD, which amounted to roughly $29,700 USD. This corroborates his estimate. 

When we asked Chow how he thinks the export-banned GPUs get into China, he stated, “So these GPUs are almost certainly moved one at a time, right? It’s very hard to get a full HGX system.”  

Later, Chow introduced us to his friend and associate Dr. Zǐ Háo Fù, who works in the university’s linguistics department and has a background in computer science. He uses the university’s mainline datacenter to train large-language models. He shared that they currently use A40 GPUs, gaming GPUs, and sometimes have an RTX 6000 available (NVIDIA has multiple generations of RTX 6000, including Ada Lovelace and Blackwell). Zǐ Háo elaborated that memory is the “most important part for researchers.”

When we asked Chow (hypothetically) if they could get whatever GPUs they wanted with limitless money, he explained that if they had millions of dollars before the GPU ban, they could get them. He also stated that, “With the export ban, [university] departments are not necessarily willing to publicly list their computing capabilities.”

Datacenter

Chow showed off the GPU cluster he constructed. He explained that a single 8-GPU system can put out around 4,000 watts. For reference, this would require 2 standard US residential circuits to support at 20A per circuit or, as Dr. Chow says, would be about “two hair dryers.”

He also shared that you can’t plug the GPUs into a standard wall socket as that would immediately blow the fuse and that they had to install 2 three-phase power within their GPU cluster room. 

Chow shared that they have roughly 30 A100-class GPUs and a bunch of 3090s and 3060s. This amounts to roughly 50-60 GPUs.

Discussing the logistics of how the US determines which GPUs should be banned, Chow stated, “If you go back like five years ago, everyone probably would think that like, yeah, FLOPs is a very important metric to consider. But turns out when it comes to loading large models, what we all care about is how much memory, how much VRAM you have, right? You simply cannot load a model if you don't have the VRAM. So now H20 becomes a very attractive option.” He added, “the H20 has drastically lower FLOPs, but then it actually has more memory than the original H100.” For reference, the H20 has 96GB of memory, whereas the H100 has 80GB. Chow added, “So it’s actually in some sense more attractive.”

While the idea of black markets might convey back-alley deals, Chow says that the export-banned GPUs are simply shipped to customers. 

When we asked if the GPU bans were effective, Chow responded, “It is effective in preventing the building of a very big cluster. It’s just not possible to get 100,000 of the GPUs.” 

He mentioned that the H200s were also assembled in China, which raises an interesting question of what prevents these GPUs from 'falling off the line.' Chow elaborated on that, stating,  'I don’t understand how it works at all. How is the ban even working?'

When we asked them why they think the US government cares so much, Zǐ Háo stated, “I guess they just want to delay the speed of other countries of training the model, but their ban is very weird.” When we asked them if the ban seemed targeted at military uses or Chinese companies in general, Zǐ Háo said, “I guess it’s in general.” Chow chimed and said, “I don't think, from the US government's perspective, there's this distinction between like Chinese academia versus Chinese military versus Chinese commercial. I don't think they really consider there's a distinction.”

When we asked to see if Huawei’s hardware seemed compelling, Zǐ Háo shared that he doesn’t know people who use Huawei hardware, but it may be a fallback option in the future if consumers in China can’t get NVIDIA GPUs.  

All these high-end GPUs are still manufactured in China, but they're not allowed to be sold in China.

- Dr. Vinci Chow

Chow explained to us that the A100 GPUs he acquired were “assembled in China.” He mentioned that the H200s were also assembled in China, which raises an interesting question of what prevents these GPUs from “falling off the line.” Chow elaborated on that, stating,  “I don’t understand how it works at all. How is the ban even working?” He points out that the box for the banned GPUs even “clearly state that they are manufactured in China.” Chow shared a theory on how banned GPUs are able to be sold in China, “My guess is there must be spares. I don't know. Spare SXM modules, spare casings, and then somehow these spare parts just get assembled into a complete GPU and get sold.”

I would be surprised if they don’t, right? I would be surprised. I would be really surprised if they don't. These are very expensive items. I would imagine you would keep track of everything

Dr. Vinci Chow

Recapping the ridiculousness of the situation, Chow stated, “All these high-end GPUs are still manufactured in China, but they're not allowed to be sold in China. And somehow the US government thinks that's going to work and somehow the Chinese government also allows that to happen. I have no idea how actually that whole thing works.”

We asked both professors if NVIDIA knows whether all of this is happening and Chow stated, “I would be surprised if they don’t, right? I would be surprised. I would be really surprised if they don't. These are very expensive items. I would imagine you would keep track of everything.” 

Hunting for Banned GPUs (Golden Computer Center)

After this discussion, Zihao parted ways and Vinci brought us to one of his favorite tech spots in Hong Kong: The Golden Computer Center and the outdoor Apliu Street tech flea market. We asked him if we could go find some supposedly banned GPUs, like the RTX 5090, just available out in public. The hope was to find a shopkeeper with some ground-level or consumer knowledge.

When we asked Chow if the export controls are having their desired effect for the US, he responded, “Well, yes, in terms of stopping China from building a comparable GPU cluster to the US.”

While we were visiting Golden Computer Center, Chow pointed out some purchasable “parallel import” GPUs, which means they were smuggled into the country. We saw banned cards for sale here, including the 5090 Founders Edition. Despite the sometimes inflated prices, we were surprised at how easy these GPUs were to purchase. It seems like one of the main impacts of the GPU bans are the prices of the cards. Chow theorizes that the cards are moved into the country one by one, often by traveling students. 

We asked a shopkeeper there, “What’s the most popular card?” He said it was the 5090 and that they come from Australia and Taiwan, neither of which has an export ban on these GPUs to China. He also said they can “buy a lot” of 5090s from mainland China. 

Hong Kong was impressive and filled with character and culture -- and GPUs, apparently, because one of our next sources sent a message the night we were planning to hitch a ferry to Shekou Port in Shenzhen. He told us that we could buy a GPU as soon as tomorrow if we wired him money immediately.

That seems like the responsible thing to do, and because it’s fully above board as a buyer, and because he texted us a photo of an RTX Pro 6000, which was very intriguing since it’s Blackwell, we decided the best way to learn more about buying export-controlled GPUs in China would be to just do it. So far, all the contacts were aware that some form of factory repurposing, theft from the line, QC rejects, and actual by-hand smuggling are involved -- but none knew for sure how the GPUs move. To get closer to the sources, we decided to wire the funds and get an address.

We ended up sending out $3,289 via wire on a tight deadline, with our boat bringing us to the seller the next morning. 

Turn a Blind Eye: How GPU Smuggling Works

HUIZHOU: Mr. 5

We then packed up and boarded a ferry to Shekou Port, but we decided to take a quick detour: Rather than go straight to Shenzhen to meet our GPU plug, we first went to meet up with someone else -- a source who told us he has more information on how GPUs get into China. We traveled to HuiZhou to meet a guy known as “Mr. 5.” 

HuiZhou is a city with some serious grit.

The city has grown to take some of Shenzhen’s factory industry as it’s been pushed out over the years with Shenzhen’s expansion and as it’s turned into more of a metropolis. We spent a good amount of time in HuiZhou over the last decade, mostly visiting case, painting, tempered glass, and tooling factories.

"we asked him how 5090s get into China. He responded, 'It’s like this. China already produces the heatsinks and components. China makes a lot of 5090s. That’s the first way in.'"

Mr. 5 is a cooling hardware reviewer and has specialist knowledge in factories that make video card cooling solutions, including NVIDIA’s. This experience allows him more access to information about the peculiar relationship between factories making cards and the companies that technically can’t sell them to the country where they’re made. We instantly related to Mr. 5 for his own editorial disputes with NVIDIA’s review sampling process, where we have felt the company seeks to control review direction. 

They can buy it and use a ‘human flesh backpack.’ They carry that back and it’s not illegal in China

- Mr. 5

His username is “51972” on Bilibili.

“Human Flesh Backpack”

Speaking to him in Chinese, we asked him how 5090s get into China. He responded:

“It’s like this: China already produces the heatsinks and components. China makes a lot of 5090s. That’s the first way in. The second is [...] The US has a ban on sales to China, but there’s no ban in China. Because Shenzen and Hong Kong are close, there’s only one customs check. Many people can get it from Hong Kong or other countries. For example, America, Japan, Singapore, etc. They can buy it and use a ‘human flesh backpack.’ They carry that back and it’s not illegal in China. They can purchase it that way. There’s one other method: for some people, it’s just for money and they choose to smuggle. So they use a ‘special channel’ (smuggling) to get it back.”

A few months ago, if you carried an RTX 5090 in from outside, you could earn about 2,000 to 5,000 RMB. You could make so much money.

- Mr. 5

We asked Mr. 5 if he thought most of the banned GPUs come in one at a time, and he replied:

“There are many ways. In China, we have a special type of ‘job,’ scalper. He can go back-and-forth many times in one day to bring them back, and every time he does, he brings some back. He can also organize people or a group of people to go to Hong Kong together to buy. Then he gets the difference in price. Each card gets from 500 to a few thousand RMB. That’s the method he uses. Strictly speaking, this kind of action is illegal, but it’s a gray area. They can organize people from Shenzhen, Hong Kong, or other countries. Of course, there are many other ways, like international students bringing them back. In China, the 5090 is not banned.” He added, “A few months ago, if you carried an RTX 5090 in from outside, you could earn about 2,000 to 5,000 RMB (about $280 to $700 USD). You could make so much money. The craziest is when the 5090 released during Chinese New Year, some people who brought one back from overseas made 10,000 RMB (about $1,400 USD). So a lot of people thought, ‘If I go abroad, I’ll buy an extra one.’” 

Of course [NVIDIA] know. For this matter, of course they know, but, how, ‘how do I say this?’ ‘Open one eye, close one eye.'

- Mr. 5

Mr. 5 stated, “China has an old saying, ‘All the hustle and bustle in the world is only for money and interest.’ I also have to add that many of the RTX 5090s are ‘made in China.’ The video card was born here locally, so a few brands choose to sell domestically in order to digest inventory. Because it takes a few months to ship by ocean freight--we already know it takes 1-2 months to get from China to the USA by boat, which wastes time. But if it’s sold domestically in China, it can be turned-around quickly. Funds return faster. This is also lower pressure and reduced inventory.”

NVIDIA’s Awareness of Smuggling

When we asked Mr. 5 if he thinks NVIDIA knows about what’s going on with all of the banned GPUs being sold in China, Mr. 5 replied:

“Of course they know. For this matter, of course they know, but, how--how do I say this?--‘Open one eye, close one eye.’” This translates to “turn a blind eye.” 

When we asked Mr. 5 if NVIDIA would want to stop it, he simply replied, “No” since the Chinese market is so big. 

SHENZHEN: Suppliers, Middlemen, & Fixers of Banned GPUs

Our next stop was in Shenzhen Bao’an to meet with a trading company that sits between Hong Kong and Shenzhen warehouses. They’re the trading company’s trading company, and we planned to buy an RTX 5090 from them. 

Shenzhen was a fishing village just 40 years ago. Now, it’s one of the most technologically advanced cities on earth. Shenzhen has some extreme Cyberpunk vibes with its mix of technology and surveillance.

It has packed tech markets in Huaqiangbei, including some of the weirdest computer parts we’ve ever seen. There are entire buildings dedicated only to small phone repair shops with their own kiosks, another building dedicated to a mix of computer hardware, gaming, and miscellaneous components, more still for just integrated circuits, and all of these places have people who know people. That’s their job. 

But with Shenzhen’s technological rise comes with it an uncomfortable omnipresence of CCTV and facial recognition, which feels more fitting today than ever before now that we know worldwide government facial recognition is a big user of AI GPUs. 

Buying an “Illegal” GPU

We started the day by meeting up with our translator for the next two days, Raymen, who’s helped us on factory tours for years now. Because Uber doesn’t work in China -- since it uses the Google Maps API and Google is blocked by the firewall -- Raymen also was our man with the locally compatible apps to get us places. This became especially important now that even cash is becoming less acceptable by some cafes and restaurants, as everyone has moved to paying with the WeChat app.

With our ride booked, we set off to see if the wired funds turned into a GPU and tried meeting up with our GPU supplier.

Upon meeting with them, we explained that we were media, and they allowed us to record them talking but didn’t allow us to show their faces or divulge their names or company name. 

They gave us the RTX 5090 we purchased, which wasn’t even much more expensive than what we would have paid for it in the US. 

We asked them what the most common GPU is that their customers buy and one of them, whom we’ll call “Silver,” stated the H800, A800, and 5090 D. These are are all banned GPUs (the 5090 D was not originally banned). We then asked them what sells better between the 5090-class cards and the H100 type GPUs, and they replied “5090 D.” 

We asked them if anyone buys high-end AMD GPUs and they responded, “Very few. AMD’s GPUs are rarely useful. We’ve had second-hand customers, but high-end is very rare.” We then asked to see if people were buying Intel GPUs, and they responded, “Intel is the least!”

'睁一只眼, 闭一只眼 '/ 'Turn a Blind Eye'

- "Silver"

We then asked them if they think NVIDIA knows about people buying and selling banned GPUs in China and they echoed what Mr. 5 said and replied, “They ‘open one eye, close one eye.’ They know you can buy it.” 

The most important thing we learned here was where to go next: Just like how the ants move GPUs piece by piece, we’ll have to collect our information piece-by-piece. The company said that some cards have become more difficult to get, but they can still get them. 

The fact that they texted us a photo of an RTX 6000 PRO Blackwell card -- the very same that we just bought for $8,500 in the US and that is hard to get at home -- shows that they’re resourceful. Their price is $8,600 US, which is actually cheaper when factoring-in US taxes and shipping. 

So the ban isn’t stopping them, and the 6000 PRO Blackwell GPU is a serious AI card (that we’ve benchmarked) with 96GB of memory. It’s banned and in demand. This company isn’t used to single-GPU sales like ours: It mostly transacts with other trading companies and in high volume, interfacing with a Hong Kong distributor to bring the cards into Shenzhen, moving it across one more border, another ant in the chain. That meant they could connect us with their distributor, which would be familiar with where we could find smugglers, but this company itself neither knew many smugglers directly nor knew many end users. That’s OK, because each link in this chain will get us one ant closer to the information we need.

We took note of their information on the distributor and smuggling side to use when we got back home, then they helped point us toward our next stop: Huaqiangbei, which is located in Shenzhen.

We hopped in a car and drove 40 minutes to meet with a GPU trader we found in the city, over near Huaqiangbei. 

SEG E-Market: The Biggest Tech Market in the World

Huaqiangbei is home to the world-famous SEG E-Market, or Saige, where we’ve found some of the strangest computer parts we’ve ever seen. Huaqiangbei has the highest concentration of technology and integrated circuits in the world, with a neverending maze of multiple disconnected, multi-story malls specializing in all electronics. If your life depended on getting a complete product made in a single city block, Shenzhen Huaqiangbei is your best bet.

Fortunately, this time, we’ll be with this guy:

This is Vincent. His profile picture on one messaging app is Van Gogh, so he has a sense of humor. He seems to have a natural ability to make people -- and cars -- get out of his way and he argued with security guards about us being able to film.

Vincent generally seems to have an attitude of getting shit done.

But before we met him, we had some concerns going into this one that it’d be fruitless or that no one would even be there to meet us. Luckily, he met up with us after we arrived. 

We can only move [GPUs] slowly, but there are a lot of people in China.

- Vincent

We asked Vincent if it was difficult to get the GPUs. He replied, “The market has a lot of them” and pointed us towards the nearby SEG E-Market. Vincent runs a trading company and buys and sells the GPUs. We asked him how the high-end GPUs get into China, and he said, "They often come through Hong Kong” and added that “Taiwanese people also sell them here.” He confirmed that the GPUs are often brought over one-by-one. When we asked him if this one-by-one movement was enough, Vincent replied, “We can only move them slowly, but there are a lot of people in China.”

If it’s a common card like a 5090, we have them in China. Chinese factories made those [5090s] and sell them

- Vincent

Consequences of Smuggling

We then proceeded to ask if there were any consequences from bringing those cards in. He replied, “If it’s a banned video card, then China doesn’t have any first-party [cards]. You can only bring it in from outside. If it’s a common card like a 5090, we have them in China. Chinese factories made those [5090s] and sell them. That’s the way it is. If you’re talking about high-tech servers, you can only bring them in from outside. China doesn’t have these [high-tech parts].”

When we asked Vincent if China cares, he responded, “It’s not the Chinese government’s business. It’s a US ban.”

He revealed that “each person has their own method to get cards in. Normally, I just get it from Hong Kong because it’s close.”

We asked Vincent what his normal order amount was and he replied, “Relatively low. Just 1 or 2 [per customer].”  We then asked what was the most in-demand GPU, and he responded, “The 4090 is relatively popular, but the 5090 is too expensive. No one wants it.” When we pointed out that the 4090 was banned, he was surprised and stated, “Really? 4090? It’s relatively common.”

We found it interesting that some dealers of "banned" GPUs aren't even aware that they are banned, illustrating just how easy these parts are to get.

Vincent stated, “We don’t know the reason for the ban, we can only buy and sell what’s on the market.”

We then asked Vincent if we could follow him to SEG E-Market and walked there with him. 

Vincent led us over to various large wholesale and consumer retail markets. His office is within about a 10-minute walk from SEG E-Market, which is his daily haunt.

There’s one peculiar detail in all of this: Vincent has customers both in and outside of China, but his foreign customer base is relatively high. He mostly deals in RTX 4090s, which seems to be a trend as you’ll see with our smuggler contact at the end.

That means that Vincent is buying 4090s that either never left China -- but were supposed to -- or were likely illegally re-exported to China, and is then re-re-exporting them to his foreign customers. It’s like an infinite loop.

He used to work as a shopkeeper in the mall itself and would interface with factories or anyone else who wandered in. 

we went into a shop and asked if they had 5090s. The shopkeeper said that they do

These days, he’s closed shop and is a buyer instead. He mostly takes orders online, doesn’t keep much inventory, and then just walks across the street to buy whatever was ordered that day. It’s basically a concierge GPU-picking service.

While at SEG market, we went into a shop and asked if they had 5090s. The shopkeeper said that they do and they confirmed that they did have them and could sell them for 19,800 RMB, which was roughly $2,757 USD. That’s not bad and is pretty close to US pricing. 

Speaking to shopkeepers, and like with the other interviews, we learned that people rarely want AMD GPUs and basically no one wants Intel.

Along the way, he introduced us to his cousin in the integrated circuit business. 

Vincent walked us around and repeatedly expressed confusion at the ban list. He sees RTX 4090s everywhere and RTX 5090s are easy to get

His cousin connected a missing link for us: People like him, selling FETs, inductors, capacitors, resistors, MLCCs, and so on for repair shops, can help to supply the parts needed for upfit or modifications. For GPUs, this can also include sourcing additional VRAM for modifications.

some sellers told us that they’re capable of sourcing devices like the A800 or even A100

Vincent walked us around and repeatedly expressed confusion at the ban list. He sees RTX 4090s everywhere and RTX 5090s are easy to get, just expensive. He was pretty sure that their abundance meant we were mistaken, but we checked and re-calculated, and they are export controlled devices. He told us about how 5090s are easy to get because the local factories supply them, matching Mr. 5’s comments earlier. Vincent pointed out everything from RTX 20-series GPUs, GTX 16 cards, 30-series, A1000 cards, Quadros, BTC mining rigs, and 5090s, and 4090s. Although some sellers told us that they’re capable of sourcing devices like the A800 or even A100, they did not have them on-site. They have them in a more secure spot. 

That’s because there’s a separate area for the warehousing. Fortunately, we learned where they are from our new friends in SEG. Like every other link in this chain, we kept getting closer to understanding the full story.

Warehouses

On a rainy morning, we headed over to the warehouses containing all of the GPUs to get a better idea for how many there might be. SEG’s guards have never been particularly friendly toward filming to begin with, but ultimately, they’re not police. The worst that happens is we get asked to leave. Besides, there weren’t any signs saying we couldn’t be there.

From the perspective of the United States, the warehouse we visited is filled with highly illegal items. They didn't have export licenses.

From the perspective of people working here, however, they’re just doing their job. It's not shady. It's not some subterranean refuge of firearms. It's just a poorly lit warehouse like any number of other warehouses around the world, but the perspective of the American government would be different from the people working here, who just want to sell these things and go home.

If GPU sellers don’t have certain GPUs, they purchase them from these warehouses. We saw stacks upon stacks of GPUs for sale. These feed into SEG and get distributed to Alibaba sellers like Vincent and get sent out from there. To be clear, not all of the GPUs we saw in the warehouse are banned, but we did see a couple 5090s. 

From the perspective of the United States, the warehouse we visited is filled with highly illegal items. They didn't have export licenses. From the perspective of people working here, however, they’re just doing their job. It's not shady. It's not some subterranean refuge of firearms. It's just a poorly lit warehouse like any number of other warehouses around the world, but the perspective of the American government would be different from the people working here, who just want to sell these things and go home.

The warehouse security did eventually kick us out for filming, although they were oddly polite about it, which we appreciated.

Zhengzhou

Next up, we headed to Zhengzhou on a 5 hour and 50 minute high-speed rail ride to visit Brother Zhang’s repair shop, where they build custom 48GB GPUs. 

Brother Zhang is famous for repairs on Bilibili. He’s got video cards and spare parts all over his shop and he saves everything that he can salvage, then rips the good components from the boards to reuse them for other ones. Brother Zhang is an interesting guy in the chain because he’s not directly part of the GPU black market -- he just fixes video cards. He does that for consumers and for companies. Sometimes, that means people want him to modify video cards to, for example, double the VRAM. That’s where it gets interesting for our story.

Speaking to him, we learned that his shop tests about 50 GPUs a day and, while we were there, we saw them testing 4090s and 5090s. Most of the GPUs are from NVIDIA, but there are some from AMD, with relatively few from Intel. Most of the cards that come into his shop come from all throughout China and most of his customers know him because of his videos.

"He converted a 24GB RTX 4090 into a 48GB RTX 4090"

The ultimate reason we’re at Zhang’s shop is to see how video cards are modified. Because if there’s an export ban on GPUs going to China, then it becomes critical for China to be more self-sufficient in keeping those GPUs that they do manage to get to stay in service. Most of the repairs are more typical things you'd expect of soldering. There's board heaters, soldering irons, surface mount components gathered from suppliers nearby, and a lot of test stations. But one thing that's unique to China and especially to this shop is the ability to take an existing model video card and completely modify the SKU into something that NVIDIA doesn't even make. NVIDIA is intentionally restrictive with how much RAM it puts on cards. Part of that is to upsell people to more expensive models. Actually, that's pretty much all of it. That's basically why they do it. But at shops like Zhang's, they double the VRAM on some of the cards that come through, making it into a fully custom SKU. We wanted to see if he had one lying around that he could show us, and he did one better than that: his shop made a modified 48GB RTX 4090 right in front of our eyes.

He converted a 24GB RTX 4090 into a 48GB RTX 4090. The unit that we looked at went through 6 repairs: It originally didn’t work at all, then it had a memory problem, then it had display issues. While the card works now and the GPU itself is functional, because its board has had so many issues, it would be better to harvest the expensive components (like the GPU) and put them on a new board to avoid other potential failures. 

When we asked Zhang what the success rate was, he told us it was about 99%. Impressive.

Because NVIDIA doesn't make a 48GB 4090, the repair shop has to source its own PCB and cooling solution. Other shops commonly sell both of these on the open market in China. It needs something that has enough pads for all the memory modules they're adding to it, more memory, and a pre-populated VRM. They work with a third-party supplier that builds a PCB, uses an SMT line to place all the VRM components, and has the extra wiring, circuitry, and pads to support 48GB of memory. 

To create a 48GB 4090, the first step is to disassemble the card.

From here, they heat the GPU to 260 degrees Celsius for about 5 minutes, which allows them to pull the GPU off the PCB. Zhang revealed to us that they get about 10-20 orders at a time and, when we asked him if NVIDIA had contacted him to tell him to stop, he responded, “I don’t think they will.”

The next step is to work on the memory. To do this, they place a PCB on a board heater and use a hot air station to heat pinpointed memory modules and remove them without risking damage to other components on the board. From there, they’ll add them to a newer board. This, ultimately, allows them to double the card’s memory capacity. 

Next, they took out a template for the solder balls and cleaned it with rubbing alcohol. From there, they poured solder balls into the template and positioned it on top of the memory modules. This allowed them to bake the new solder balls onto the memory modules, which they could then attach to the PCB. They then applied flux to wet the solder balls and used a solder wick to pull off the excess solder.

From start to finish, it took [the shop] about 2 hours to make a 48GB 4090

When we asked Zhang what 48GB RTX 4090 cards normally sell for, he told us over 20,000 RMB, which is roughly equivalent to $2,785 USD. That’s not a bad price.  

From there, he put a lid on the template and poured in soldering balls, allowing them to adhere to the memory. They then heated the modules on a hot plate that ran at around 195 degrees Celsius. Because the template only allowed them to do 8 at a time, they had to do the solder ball process twice. They then applied heat directly to the surface where the solder balls were using a hot air station.

The next step involved cleaning the GPU in order to mount it to a new PCB. From there, they added flux to the GPU. They then applied a custom template to the bottom of the GPU and poured solder balls on top of it and used the template to sift the solder balls.  

The next thing the shop did was place the GPU on top of a jig to hold it in place on top of a hot plate to protect it.  

They then placed the memory modules onto the PCB and, again, used the hot air station.

Finally, after about 2 hours, the shop added the GPU to the PCB and placed it under heat to bake the GPU back on.  

From there, the team cooled the GPU off with a fan and added thermal paste. A technician placed a heat sink on top to test the card to see if it worked, leading to an instant success.

From there, they attached the GPU and PCB to the rest of the video card’s cooler and chassis. From start to finish, it took them about 2 hours to make a 48GB 4090. 

At the beginning of this process, the card was working, but it had been through so many repairs that it was very likely to break again at some point from something they couldn't really predict. From a waste standpoint, it's just better to figure out a way to save that card because the PCB is kind of the least valuable part and is also the least precious of the resources. Being able to save the silicon that's in the memory and in the actual GPU itself provides a lot of value, but it especially offers value in China, where they just don't have as much supply, even though the supply looked good from what we saw. Purely from the perspective of people in China, not only is this better than what NVIDIA shipped to begin with, and really not that far off in price when the card was at its most scalped, it's also just a good way to keep the silicon in circulation even as things like MOSFETs, capacitors, or PCBs die.

The silicon tends to be pretty resilient and it doesn't really die that often on GPUs. What this tells us is that this is maybe an alternative method to getting more GPUs into supply. If they take the broken boards and then put functional silicon back into circulation, it's a certain level of ingenuity, and there aren't many places that are set up to handle it. Brother Zhang's shop is impressive.

Video Card Factory

Next, we headed back to Shenzhenbei in a high-speed rail trip that added roughly 1,000 miles to our journey, planning to meet up with Yeston Creative Director of Product Zhou. Yeston is a video card manufacturer.

Zhou and Yeston were not officially contributing to the black market side of our research, but they did show us around their factory and there were a few interesting things we learned that relate to our story. The big takeaway is that factories receive their GPU supply from their partners. That would be AMD, Intel, or NVIDIA, and the memory is often packaged with it. But for the most part, everything else that they use, unless they're buying reference PCBs, all come from whatever sources they want to find for their supply. 

for at least the gaming class cards, which the 5090 more or less is one of, there's enough volume and NVIDIA’s hands off enough that it would be pretty easy to make these disappear

We were also informed that there's very little oversight in terms of the management of the rejects at these places. They can file for refunds if, for example, a chip is bad. But if they screw up a board in their own process, then it's going to be on them to deal with the defect.

Factory work for GPUs is highly automated. For at least the gaming class cards, which the 5090 (more or less) is one, there's enough volume and NVIDIA is hands-off enough that it would be pretty easy to make these disappear. The biggest reason for this is that there are partners for the 5090s, unlike cards that only NVIDIA makes, such as the RTX Pro 6000. Since there are partners, NVIDIA is only tracking when they sell those partners the GPUs. NVIDIA doesn't necessarily keep tabs on where those get distributed once they're done being manufactured. For this reason, if a factory wants to just sell the GPU domestically in China and they think they can get away with it without being added to the entity list, which would very likely kill their business with NVIDIA, then they might sell them domestically. That might be how some of these trading companies get a hold of them. It's an easier way to make them disappear than through channels that are more publicly visible from NVIDIA and the government. 

Yeston doesn't make 5090s, but the processes they taught us about illustrates how that could happen. 

As for the server grade solutions, there are manufacturers for those that are pretty tightly controlled. Basically, the silicon is made in Taiwan, shipped to wherever it's going to be assembled, and accompanied by the other components sourced in China. As a result, it's easiest to keep the whole supply chain there, including the assembly. 

after seeing how a highly skilled group like Brother Zhang's repair shop can just repurpose all those components that are still good, you could see how there would be value in figuring out a way to make that board disappear, even if it's supposed to be in the trash

Factories use surface mount technology (SMT) lines with conveyed inflow of PCBs and reels upon reels of components that are placed smallest to largest, with the most valuable typically at the very end. That would often be the GPU and VRAM. Heatsinks are installed often through a manual process at the end. There is often manual assembly even for NVIDIA's high-end products.

We have seen NVIDIA's server video cards (and servers) being made in different factories. If the factory finds a defect, however unlikely, it could 'disappear' into scrap to be resold to the domestic market later. Dr. Vinci Chow had such a board where he had a broken link, and so there's a good chance that card stayed in China and never left. If it was reported as a defect to NVIDIA, then it may have effectively just been written off and considered trash. After seeing how a highly skilled group like Brother Zhang's repair shop can just repurpose all those components that are still good, you could see how there would be value in figuring out a way to make defects -- or "defects" -- disappear, even if it's supposed to be in the trash. 

This comes back to NVIDIA turning a blind eye, because they'd eventually notice it from the serial number popping up somewhere. NVIDIA might choose not to notice it, though. Besides, once someone has a card, especially in China, there's not a ton NVIDIA can do, anyway.

Taiwan

We then headed to Taipei, Taiwan. Most of our stay in Taiwan was for another story; however, we did accidentally unturn one stone with something interesting under it.

A B2B company in Taiwan noted to us that it is commissioned by Chinese companies to import servers for pre-testing and pre-assembly and setup steps. When we asked what GPUs the servers ran, we realized that all of the hardware is export-controlled.

Once the B2B agency completes its testing, it ships the system out to the original buyer. Basically, they act as an intermediary to buy the machine, then mark it up and reship it to the original purchaser.

We also spoke with a company that conducts business in Singapore. The company informed us that they are also aware of similar passthrough so-called “testing” services.

Smuggler

Our last link in this chain is actually the first: The smugglers themselves, and we’re back in the US for this. We didn’t think we’d be able to find one doing the dirty work, but in the final hours producing this story, one of our viewers provided a lead. This led to about a day of production delay as we were wrapping this project up, but as far as we’re aware, this is the first content piece that actually features someone doing the highest risk dirty work.

Everyone else we spoke to is in China. They’re safe from US retaliation. But large-scale smugglers get arrested, fined millions of dollars, and can spend years in prison if they were serious volume movers.

A viewer contacted us to say that he’d connected with a traveling GPU buyer. The story went that this guy drives around buying specifically RTX 4090s. He doesn’t care about anything else, including 5090s, because those don’t sell as well to China. 4090s are ideal, he told us, because they can be modified into cost-effective 48GB models, like at Brother Zhang’s shop.

The smuggler was extremely open with us in text messages written only in Chinese, but wasn’t open to us flying out to him. He did allow us to share footage of his car, however.

The Plug offers $2,000 flat per RTX 4090, which isn’t a bad price. He told us that he then finds ways to get them into China

The Plug, as we call him, has an ATX test bench, motherboard, simple downdraft cooler, power supply, obviously a CPU, RAM, and an SSD, and gigantic battery in the trunk of his Prius. He also has at least one spare license plate in his trunk in addition to the one on the back of his car. We’re not sure why, but it’s better not to ask.

People like The Plug post on Facebook Marketplace and other common online reseller forums looking to buy GPUs, just like anyone else in your city would. The Plug offers $2,000 flat per RTX 4090, which isn’t a bad price. He told us that he then finds ways to get them into China. He recently was scammed out of $5,500 of payments owed by a Hong Kong buyer that resells in Hong Kong, so lately, he’s been considering hand-carrying the devices in luggage instead.

His margin is slim. He makes just under $300 US per GPU, not counting gas, potential hotels as he travels, and time. Our understanding is that some in his shoes will strip the GPU cooler off and ship just the PCB back to China, increasing their margin. Coolers are available in abundance where they’re made, so it can be cheaper to do this than to pay for the weight and size. 

He sometimes buys entire computers just to take the 4090s out, as that’s what the bounty is on. He then sells the remaining system back to anyone who’ll buy it. He told us that 5090 prices in China are falling so fast due to oversupply, ironically, that he’d lose money reselling it to China and would do better flipping it to Americans.

The Plug isn’t rich. He seems to be doing OK, but it’s not the type of wealth you might expect for such risk. 

We can’t overstate how important it was to get this piece into the story, as this allowed us to fully complete the chain. We want to provide a huge thanks to our anonymous viewer for their help. 

Many of the sources we met in Asia during this trip told us they simply didn’t know how the GPUs actually get in, and now we know. People doing this on a small scale like The Plug are unlikely to be caught, but operations transacting millions of dollars worth of GPUs would have a harder time getting them out undetected.

That brings us to the end of our travels, but not the end of this piece. We still have two more entities to look at, and that’s NVIDIA and the US Government.

Details: Legality

The legality is simple -- this is our understanding of it. 

There is no restriction on purchasers, only on sellers the US has control over. Even Americans buying GPUs in China are not violating any laws, as the purchase of a GPU in China is not restricted.

We’ll start with selling:

The simplest answer to “who the sale of GPUs is restricted for” is anybody who doesn’t have a specific export license and who would be governed by US law, and NVIDIA can’t just bypass it by shipping from a different country. It has to do with the sale, not with the shipment location. That includes entities buying from US companies, like European companies, who would be under guidance both from NVIDIA and, if they do business in the US, its government. 

It’s also illegal for anybody in the United States, citizen or not, to sell these GPUs to China, Hong Kong, Macau, or companies in those locations if the seller does not have a re-export license.

It is not illegal for a person of any nationality in America to sell a GPU to any non-restricted entity.

As for buying, as Dr. Vinci Chow stated, “There’s absolutely nothing illegal for us over here to buy these GPUs.” 

But it’s also not illegal to sell them in China for a Chinese company. The Chinese government doesn’t generally enforce American export laws. Other nations cooperating with the United States might, such as the recent arrests in Singapore, but once the GPU is in China, the people in possession of it likely don’t care -- buyer or seller or someone who does both, like GPU dealer Vincent.

The only control over Chinese companies that the US has is the Entity List, which would hurt their business prospects with American companies -- but only if they care about that. Chinese GPU middleman 思騰合力科技有限公司 (Sitonholy Technology Company) in Tianjin landed on the Entity List for transacting banned GPUs and being found out. Sitonholy purchases GPUs from anyone who can get them into China, and then they bid on domestic projects like data center build-outs.

The Entity List was used to restrict, for example, DeepCool previously, resulting in their American partners ceasing business with them for fear of frozen assets, audits, or collateral bans. This hurts companies like DeepCool that want to operate in the US and even shut down their California office, but for Sitonholy, they mostly want to do business with other Chinese companies, and so it’d have limited impact unless they wanted to expand to do business with Americans.

Beyond the entity list, the US really has no control over what happens inside Chinese borders. That means the only point at which a GPU could feasibly be intercepted and a person arrested would be operating in the US or in transit to intermediary countries, such as Singapore, which may have their own export laws. The people buying and selling them within China are not breaking any of their own government’s laws, though.

We think the story of AI GPUs has become a story of corruption between governments, and the wealthiest company on Earth.

NVIDIA is Playing All Sides

And so we come back to NVIDIA. At every turn, it really looks like NVIDIA is playing all sides. If there’s enough money to be made, anybody is NVIDIA’s friend. 

We think the story of AI GPUs has become a story of corruption between governments and the wealthiest company on Earth.

For example, on April 30th, Amazon-backed AI startup Anthropic called on the US government to increase export control restrictions to China. As part of a blog post, Anthropic said the government needs to improve its export enforcement to reduce smuggling. The company cited examples of chips being smuggled with “prosthetic baby bumps” and “live lobsters.” This upset NVIDIA, obviously, because NVIDIA doesn’t like restrictions on making money.

So, NVIDIA shot back. Seemingly taking a page out of Trump’s playbook, NVIDIA essentially called this fake news and stated, “American firms should focus on innovation and rise to the challenge, rather than tell tall tales that large, heavy, and sensitive electronics are somehow smuggled in ‘baby bumps’ or ‘alongside live lobsters.’”

We unearthed this official Hong Kong customs website detailing a white van busted driving on the Zhuhai-Macau bridge, filled with 280 kg of undeclared live lobsters and, yes, 70 smuggled GPUs

Gaslighting by NVIDIA

Anthropic isn’t telling tall tales--it’s right. A 2022 video previously showed a security check at Zhuhai port in Guangdong, not far from Shenzhen, wherein a woman with a prosthetic baby bump was shown to have been carrying CPUs and iPhones instead. The report made it to customs.gov.cn, stating that she arrived from Macau, a common go-between (similar to Hong Kong). This story got international attention in technical media and we reject the possibility that NVIDIA wasn’t aware of it. We found the official Chinese Government posting about smuggling from Macau, so there's public record. Imports to most of China are taxed, and so tax evasion coupled with smuggling will increase margin on the electronics rather than sharing it with one of the two governments. If you’re already breaking a US law, it seems some just go for a hat-trick and increase the profits. 

As for the lobsters, that’s real, too: We unearthed this official Hong Kong customs website detailing a white van busted driving on the Zhuhai-Macau bridge, filled with 280 kg of undeclared live lobsters and, yes, 70 smuggled GPUs, complete with photo evidence. 

NVIDIA’s response to call these 'tall tales' then is not only defensive, but serves to gaslight and grossly mislead, we think

The Hong Kong customs itself calls this “zousi,” or smuggling, and notes a maximum sentence of 7 years.

NVIDIA’s response to call these “tall tales” then is not only defensive, but serves to gaslight and grossly mislead, we think, and is tantamount to lying for sake of downplaying reality for its own benefit. But then this is a common NVIDIA tactic, including its dishonest approach to reviews that we’ve already detailed and its deceptive and we think false advertising of the RTX 5070 as being equivalent to an RTX 4090, which is provably and comically false.

Another instance of NVIDIA’s “fake news” defense was following a July 24 publication by the Financial Times, reporting that more than $1B worth of NVIDIA’s AI chips had been smuggled to China. In response, NVIDIA, whose blind eye is turning an awful lot lately, downplayed the issue, and stated, “Trying to cobble together datacenters from smuggled products is a losing proposition, both technically and economically. Datacenters require service and support, which we provide only to authorized NVIDIA products.”

That sounds like something a company selling support would say. And the first part doesn’t really match the whole “the more you buy, the more you save” assertion. It’s only a winning proposition by all of their prior years of statements. If your only option is a useless insufficient data center or a cobbled-together sufficient data center, then a cobbled-together one is still a winning proposition by comparison. It’s weird for the company to pretend that this isn’t worth doing. It’s worth lots of money. 

Tech Exec Sycophancy

That same day, whitehouse.gov posted an article titled, “Wide Acclaim for President Trump’s Visionary AI Action Plan.” Near the top of the post, it highlighted a sycophantic quote from Huang, which read, “America’s unique advantage that no country could possibly have is President Trump.” And we could call that sycophantic if it was about any president. 

On August 5, NVIDIA got another opportunity to talk. The US Department of Justice announced it had arrested two people in California for smuggling “tens of millions of dollars’ worth of sensitive microchips used in artificial intelligence (AI) applications” to China. The BBC reported that court documents say the shipments included the NVIDIA H100 and RTX 4090. Rather than admit smuggling exists, NVIDIA downplayed the situation and stated, “This case demonstrates that smuggling is a nonstarter.”

Except that anyone who made tens of millions of dollars before getting caught had a pretty good start, and so did their customers. This is a bullshit statement from NVIDIA that seemingly aims to downplay and deflect to reduce lawmaker attention on its monopoly.

NVIDIA also said, “We primarily sell our products to well-known partners, including OEMs, who help us ensure that all sales comply with U.S. export control rules.” NVIDIA noted that, “any diverted products would have no service, support, or updates.” 

Again, this is not fully true. Our own sources in this story noted that, although something like an HGX system would be hard to service, a standalone PCIe GPU could be parted-out and covered under a separate warranty, even in China. 

a government committee called Deepseek a 'threat to national security' and said Deepseek had used NVIDIA’s technology

Fears of Deepseek

Meanwhile, the US began scrutinizing NVIDIA’s technology for getting into China, whether or not the company itself was directly involved. In a bipartisan report called “Deepseek Unmasked,” a government committee called Deepseek a “threat to national security” and said Deepseek had used NVIDIA’s technology, “AI model appears to be powered by advanced chips provided by American semiconductor giant NVIDIA and reportedly utilizes tens of thousands of chips that are currently restricted from export to the PRC.” It continued, “NVIDIA designed and manufactured many of these chips to create the most sophisticated possible chip while skirting U.S. export controls. This has allowed these chips to be exported to China as the U.S. government develops stricter restrictions. Since March 2024, it is estimated that NVIDIA has produced over 1 million chips for the Chinese market.”

Singapore Suspicions

The government also examined NVIDIA’s significant revenue growth in Singapore compared to China based on SEC filings, particularly in years featuring restrictions. The government questioned “whether PRC customers are arranging for the diversion of sensitive chips that are reportedly sold through Singapore,” since revenue from Singapore had grown from almost nothing since 2021. 

NVIDIA has defended its sales to Singapore by saying that “Customers use Singapore to centralize invoicing while our products are almost always shipped elsewhere.” According to NVIDIA, shipments destined to Singapore were only 2% of the company's total revenue in 2025. But we also know that Singapore has made numerous arrests relating to GPU smuggling, so there appears to be some reason for the concerns, whether or not NVIDIA itself wants to turn a blind eye to it.

Ignorance is Bliss

Despite the allegations, NVIDIA downplayed any smuggling of AI chips. 

In a video uploaded 2 months ago, Jensen Huang spoke on smuggling, and stated, “Governments understand that diversion is not allowed. And there's no evidence of any AI chip diversion.” Except that there is evidence of it -- not only in this very story, but in readily available reports online for years now.

Huang continued, “Our data center GPUs are massive. These are massive systems. The Grace Blackwell system is nearly two tons. And so you're not going to be putting that in your pocket or your backpack anytime soon. And so these systems are fairly easy to keep track of, but the important thing is that the countries and the companies that we sell to recognize that diversion is not allowed and everybody would like to continue to buy NVIDIA technology. And so, they monitor themselves very carefully and they're quite careful about that.”

This one is interesting. Huang is right that it’s much harder to smuggle Grace-Blackwell or Hopper HGX-class complete systems. Dr. Vinci Chow’s statements align with this when he said, “It's very hard to get like a full HGX system.” But it still happens. At least one of NVIDIA’s GPU and server customers in another country told us that they facilitate intermediary transmission to China and in fact showed us the server racks on-site in their facilities. We weren’t allowed to film them, but we saw them. A separate representative told us that document forgery through third-party countries can also disguise such transshipments.

One middleman told us that an NVIDIA distributor gets parts into China; a downchain factory told us that NVIDIA’s QC rejects sometimes end up repurposed and kept in China, salvaging the GPU and VRAM and scrapping the rest; Dr. Vinci Chow told us that one of his own devices had a defective link on it, contributing to this statement. And when he asked him whether he thinks NVIDIA knows all of this is happening, he replied, “I would be surprised if they don’t, right. I would be surprised. I would be really surprised if they don’t. These are very expensive items. I would imagine you would keep track of everything, right? It’s hard to know what [a person] plans to do with all of these defective parts, but I’ll be very surprised that no one has ever thought of the possibility that, if it’s something so valuable, someone would come up with a use [for] even a defective one.”   

NVIDIA’s Hypocrisy

So in one set of statements, NVIDIA said that smuggling doesn’t really happen because the export controls work and keep partners in-line.

But in another statement, Huang called the US export controls a “failure,” talking out of both sides of his mouth, we think. He spoke of competing Chinese GPU brands posing a threat to NVIDIA, “The local companies are very, very talented and very determined, and the export control gave them the spirit, the energy and the government support to accelerate their development. I think, all in all, the export control was a failure.”

But the stakes, and dollar signs, for NVIDIA had increased. Jensen said NVIDIA’s market share in China had dropped from 95% to 50%, and, in NVIDIA’s May quarterly earnings before the H20 exemption and revenue share, Jensen Huang said the company’s data center business in China was done, “However, the $50 billion China market is effectively closed to U.S. industry. The H20 export ban ended our Hopper data center business in China.” Or, as Jensen said, the China market is worth one Boeing. Boeing is probably not the best example...

Conclusion

The October 7, 2022 Biden Administration export controls had a stated goal to 'protect US national security and foreign policy interests' by implementing new export controls

Let’s look back at the timeline once more between NVIDIA and the US government, highlighting NVIDIA’s relentless appetite for global dominance.

The October 7, 2022 Biden Administration export controls had a stated goal to “protect US national security and foreign policy interests” by implementing new export controls restricting China’s ability to build high-end semiconductors, including for the development of supercomputers. There was an included goal of staving off the potential for China to develop “nuclear weapons and other military technologies.” 

Blocking the H100 and A100 led to NVIDIA creating an export-compliant A800 at about 70% of the speed of an A100 for the Chinese market. 

A year later on October 17, 2023, the US Department of Commerce updated its export compliance and restricted NVIDIA’s A800 chip as well along with the newer China-targeted H800.

Weeks later on December 6, 2023, NVIDIA told reporters in Singapore that it would be working on another new chip that would comply with the US’ new restrictions. 

The AI Diffusion Rule

On January 13, 2025, which was after Donald Trump won the 2024 presidential election, but a week before he took office, the outgoing Biden administration tightened export controls by introducing national chip caps for many countries, except for 18 allies. That’s the AI Diffusion Rule we spoke of earlier and would have gone into effect in May.

NVIDIA, who had remained relatively quiet about the regulations up until this point, criticized the restriction and made an attempt to appeal to the president-elect, and reportedly stated, “It makes no sense for the Biden White House to control everyday datacenter computers and technology that is already in gaming PCs worldwide, disguised as an anti-China move. The extreme ‘country cap’ policy will affect mainstream computers in countries around the world, doing nothing to promote national security but rather pushing the world to alternative technologies. AI is mainstream computing – ubiquitous and essential as electricity. This last-minute Biden Administration policy would be a legacy that will be criticized by U.S. industry and the global community.” It seems like NVIDIA tried to set up an appeal to the president-elect, stating, “We would encourage President Biden to not preempt incoming President Trump by enacting a policy that will only harm the U.S. economy, set America back, and play into the hands of U.S. adversaries.”

We already went over the million dollar dinner and ensuing ban-then-unban of the H20 chip.

NVIDIA Persuades Trump Administration

roughly 1 month after Jensen reportedly spent $1 million to eat dinner at Trump’s Mar-A-Lago estate, the US Department of Commerce confirmed that it will not implement the AI Diffusion Rule that NVIDIA campaigned against

Later that month on April 30, Huang said this of Trump, stating, “Without the president's leadership, his policies, his support, and very importantly, his strong encouragement[…] frankly, manufacturing in the United States wouldn't have accelerated to this pace.”

On May 7, roughly 1 month after Jensen reportedly spent $1 million to eat dinner at Trump’s Mar-A-Lago estate, the US Department of Commerce confirmed that it will not implement the AI Diffusion Rule that NVIDIA campaigned against and that was created under the Biden administration. The rule was supposed to go into effect a week later on May 15. This does not unban GPUs like the H100, 5090, B100, and so on. 

Following the termination of what was supposed to be a rule to address national security implications, the Department of Commerce, now under President Trump, stated, “The Biden AI rule is overly complex, overly bureaucratic, and would stymie American innovation. We will be replacing it with a much simpler rule that unleashes American innovation and ensures American AI dominance.” 

NVIDIA predictably celebrated the statement. The company, which has been begging to sell to China while also praising Taiwan’s importance, now took an America-first posture, collecting countries like Pokemon, writing, “We welcome the Administration’s leadership and new direction on AI policy. With the AI Diffusion Rule revoked, America will have a once-in-a-generation opportunity to lead the next industrial revolution and create high-paying U.S. jobs, build new U.S.-supplied infrastructure, and alleviate the trade deficit.” Job creation promises coming from this company, in particular, are interesting, but they are playing all sides consistently.

Later that month on May 28, Huang spoke with Mad Money host Jim Cramer. He stated, “When [Trump] rescinded the AI Diffusion Rule, it was a visionary move. It was a bold move, and he recognizes that there’s an AI race and we’re not alone. And he wants America to win.” 

A Manipulative NVIDIA

June 23, via Reuters, an official of the US State Department, which didn’t reply to GamersNexus except the Department did send us 3 out-of-office auto responders, warned of DeepSeek military and intelligence operations and warned of the use of “shell companies” in Southeast Asia to circumvent export restrictions. The report mentioned that DeepSeek had “large volumes” of high-end H100 chips, which are banned in China. 

NVIDIA didn’t like that, responded to Reuters, and stated, “We do not support parties that have violated U.S. export controls or are on the U.S. entity lists,” adding, “With the current export controls, we are effectively out of the China data center market, which is now served only by competitors such as Huawei.” Turning a blind eye to the situation. NVIDIA added, “Our review indicates that DeepSeek used lawfully acquired H800 products, not H100.”

3 days later on June 26, The Information reported that DeepSeek’s next AI model has been delayed due to a shortage of NVIDIA AI GPUs in China. This directly contradicts Huang’s comments that export controls do not work.

On July 4, Bloomberg reported that the Department of Commerce, which also did not reply to GamersNexus’ emails, was preparing new export controls on Malaysia and Thailand to reduce chip smuggling; interestingly, Singapore, which now comprises a significant portion of NVIDIA’s revenue, was not on that list despite being a known smuggling passthrough.

[Jensen Huang] really pulled off something few tech CEOs have managed. He played both Washington and Beijing and he won

- CNBC Business News anchor Deidra Bosa

On July 10, Bloomberg reported that Huang and Trump were scheduled to meet again ahead of the CEO’s planned trip to China. Days later on July 14, NVIDIA confirmed that it will resume sales of H20 chips to China with Huang stating, “The U.S. government has assured NVIDIA that licenses will be granted, and NVIDIA hopes to start deliveries soon.” 

On July 15, CNBC Business News anchor Deidra Bosa gave her synopsis on the situation and said, “[Jensen Huang] really pulled off something few tech CEOs have managed. He played both Washington and Beijing and he won.” She added, “Jensen has stayed disciplined and diplomatic with a clear message, and that is: NVIDIA’s dominance serves America’s interest.”

We write to express our deep concern over the recent decision to resume exports of NVIDIA’s H20 chips to China. As policymakers and professionals with a background in national security policy, we believe this move represents a strategic misstep that endangers the United States’ economic and military edge in artificial intelligence.

- national security experts to US Secretary of Commerce Howard Lutnick

On July 28, The Financial Times reported that the US Commerce Department was not going to make “tough moves” to tighten export controls to China. This is in spite of several congressional members warning the administration not to loosen the US’ export controls for AI GPUs. Several national security experts also voiced their concern by sending a letter to the US Commerce Department, which read, “We write to express our deep concern over the recent decision to resume exports of NVIDIA’s H20 chips to China. As policymakers and professionals with a background in national security policy, we believe this move represents a strategic misstep that endangers the United States’ economic and military edge in artificial intelligence.”

NVIDIA might have started off as a much more humble company, but it has become a savvy political player in a game that’s seemingly pay-to-win

On August 11, via Bloomberg, Trump said he was open to allowing NVIDIA to sell modified versions of the company’s newest Blackwell chips to China.

That brings us to today. 

There’s no one better equipped to play that game than the most valuable company by market cap in the world

Most Valuable Company by Market Cap in the World

NVIDIA might have started off as a much more humble company, but it has become a savvy political player in a game that’s seemingly pay-to-win. That seems only fitting for a gaming company to be particularly good at pay-to-win games. There’s no one better equipped to play that game than the most valuable company by market cap in the world, now at $4 trillion, led by a man whose net worth is estimated at $148.1 billion. NVIDIA knows when to bite its tongue and how to effectively appeal to the ego of politicians of all parties and all countries.

We think NVIDIA is playing all sides. We think it is greedy, manipulative, and carefully employs propaganda such as its use of the “fake news” playbook for news which is literally reality. But we don’t think NVIDIA has a particular set of beliefs beyond just making more money. We think NVIDIA will sell anyone out to make a buck.

NVIDIA is in the big leagues now. Inside of one month, reportedly paying $1 million to a sitting President after which followed the unlock of $5.5B of lost H20 revenue, followed next by a 15% split of that unlock going to the US Government, is what raises these new questions of NVIDIA's integrity in our piece. 

As for the actual black market side of it and smuggling, it was an exciting story to cover and get to the bottom of. We've learned that common methods include factory so-called QC defects, hand-carried items by students, actual smugglers on the ground in the US, and suppliers through third-party countries, among others.
We loved working on this story and meeting all of these unique people. Each person played a key role in helping us find the next person, allowing us to complete the first public, on-record, complete start-to-finish cataloguing of a smuggling pipeline for high-end silicon. We want to thank everyone who made this story possible, including our viewers who funded it.

This article goes through the pricing updates for NVIDIA's RTX 50 series GPUs, AMD's RX 9070 XT and 9070 GPUs, and Intel's B580 and B570 video cards

The Highlights

• GPU MSRP is still too high in a lot of cases, but it's certainly better than the huge price overhead on top of MSRP back in June
• The rumor is that NVIDIA will stop bundling memory with GPUs to partners, which we worry will result in worse partner pricing for board vendors and higher prices
• GPU pricing has cratered since June and March of this year, and it's just in time for DRAM shortages and VRAM prices to go up

Table of Contents

• AutoTOC

Intro

The GPU Market has been in flux all year. Right when memory prices are skyrocketing, GPUs are finally available near MSRP.

Editor's note: This was originally published on November 20, 2025 as a video. This content has been adapted to written format for this article and is unchanged from the original publication.

Credits

Host, Writing

Steve Burke

Video Editing

Tim Phetdara

Research, Writing

Tannen Williams

Writing, Web Editing

Jimmy Thang

The new rumor is that NVIDIA will no longer bundle memory with GPUs, meaning partners will be on their own to source memory. That’ll almost definitely result in worse pricing for the board partners, ending up with more of an FE advantage while also fully embracing the memory supply impact on VRAM pricing.

Which, of course, is also caused in part by NVIDIA in its thirst for AI. 

We talked about the memory side in a separate story, it’s just that VRAM will ultimately be impacted by this, which will impact GPU prices.

Even still, we’re looking at GPU prices as they stand right now to update our GPU pricing series from earlier this year.

Based on the data, people still aren’t buying 8GB GPUs, and the higher end cards are still the worst examples of AVG Price and AVG% over MSRP. 

Pricing and availability in the GPU market have noticeably improved since we last collected the data in early June. 

Instead of only 13 current gen GPU models available for MSRP, there’s now 58 listings in stock at MSRP. And the number of total in-stock listings increased from 135 to 171. So far, these are good things.

Prices have also improved, with AVG price for in-stock listings decreasing for nearly every current gen GPU.

The 5090 (read our review) was also the only GPU not available for MSRP. It was also the only GPU that increased in AVG price since we collected them last, continuing to demonstrate how NVIDIA’s lack of competition from AMD and Intel in the high-end market segment inevitably leads to consequences for consumers.

So GPU pricing is getting better, generally, but it still has a lot of room for improvement, especially with the higher-tier graphics cards.

Our primary focus will be on the 9070 XT (read our review) since we have the most extensive data for the 9070 XT, and because AMD was so off its MSRP mark at launch.

Overview

In light of the ongoing price surges we’ve seen in the DRAM industry recently, we decided to revisit prices in the GPU market. 

We collected prices at a few points over the last few months, including just before Thanksgiving, when prices are in flux more often. We aren’t factoring-in Black Friday sales.

We think our data’s representative of the most stable prices we can expect to see before they’re affected by the incoming DRAM price hikes that’ll soon get passed on to GPU customers through the cost of VRAM.

We’ve already seen rumors that NVIDIA’s 50 series Super launch will be delayed as a result, and AMD seems to be following a similar path.

Found via Dan Nystedt on Twitter: “AMD has notified supply chain partners it will raise graphics card prices 10% across the entire product line due to rising memory chip prices, media report. It will reportedly be AMD’s 2nd such price increase.”

Further, the newest information suggests NVIDIA will stop selling packaged memory with its GPUs, forcing board partners to source their own memory supply -- and it’ll likely be at worse rates than what NVIDIA gets.

Today’s article will be relatively straightforward, especially compared to some of the more in-depth technical pieces we’ve published recently, like our recent Linux GPU benchmarking. 

We’ll start by examining how GPU pricing and availability has changed within the 5 and a half months since we last collected prices.

Then, we’ll take a closer look at the 9070 XT’s price history, as that’s currently our most comprehensive dataset for any individual GPUs.

And finally, we’ll wrap up with any additional observations, some extra charts we put together when trying to figure out the best ways to visualize the data, and our general outlook on the current state of the new GPU market.

But before getting into any of the data, we first need to define our methodology.

Methodology

The last time we collected GPU prices in June, we used listings from Newegg, Amazon, and Micro Center, but we’ve excluded Amazon and Micro Center listings today, instead relying only on Newegg first-party listings. Our primary reasoning for this is that Newegg has been more price-stable than Amazon, while Micro Center often integrates bundles that affect price in ways that are difficult to account for (and often in-person only). Amazon gets flooded with third-party listings that complicate things, while Newegg has the most first-party shipped & sold listings for us to work with.

We’re also working with only in-stock listings. This approach isn’t perfect and average prices will lean higher because typically GPUs sell out of their cheapest models first.

Remember also that we don’t have the actual inventory unit counts for any of the GPU models we’ll be going over today. We can use the number of models a certain GPU has in stock to draw inferences from, but that obviously doesn’t tell us the quantity of cards.

Getting into the results:

Our first charts will take a look at how AVG prices for in-stock listings have changed since we last collected the data in early June.

We also understand that other regions have different prices. For this piece, we’re focusing on the market we know the best.

In-Stock AVG Price Comparison | June to Current | GamersNexus

Compared to our June check-in, all current generation graphics cards had at least some price drop – except for the 5090.

The largest price reductions since June come from the RX 9070 XT, RTX 5080, and 5070 Ti. The 9070 XT dropped from $837 in June to about $655, or over 20%, with the $600 MSRP a further 8% below the average. $837 in June is absolutely insane for this video card. 

Since June, the 5080’s in-stock AVG decreased by $155, or 10%, to its $1,360 in-stock AVG now. The 5070 Ti followed similarly: Its in-stock AVG decreased by $130, or 14%, to its current in-stock AVG of $820.

The 5060 Ti (read our review), 5060 (read our review), and 9060 XT (read our review) are the cheapest and started closer to MSRP, so it makes sense that they were least impacted in terms of absolute reductions to their in-stock AVG prices. They saw reductions from $15 - $50. 

Intel has improved dramatically, dropping from $344 on average to $264 for the B580 GPU. The B570 dropped from $308 to its $220 MSRP. These are significant changes by percentage, at 23% reduction for the B580 and 29% reduced price for the B570.

The RTX 5090 was the only card to increase in price, with an AVG $50 greater than it was in June. Considering it doesn’t have competition, this price stability makes sense, but it’s obviously discouraging to see an increase. The 5090 will also be the most impacted by VRAM prices due to its capacity, so without price suppression from NVIDIA’s direct sourcing, it could climb the most.

In-Stock AVG % Over MSRP | June to Current | GamersNexus

Using only in-stock listings, this chart shows the average price percent over MSRP baseline.

The Arc B570 (read our review) and B580 (read our review) GPUs had outlier overages beyond MSRP previously and stand out as having the greatest return to normal pricing. 

We dug into this further and found these reductions were most notably caused by Weeliao’s Battlemage cards disappearing from listings. These were previously the most expensive listings offered for Arc models and, since Arc has so few models available, they were often among the few in stock.

MSRP for the 5060 Ti 8GB is currently higher than the retail price by a thin percentage, aligning with the fact that these are still having the most trouble selling to DIY enthusiasts. That’s probably of no concern to NVIDIA though, since SIs and OEMs will still be able to use them to mislead customers.

The 9060 XT 8GB is in a similar situation, at 2.9% over MSRP now for 8GB. The 16GB models for these cards are at 7.8% and about 8.3% over MSRP. The 5060 non-Ti is also close to baseline, at 3.5% over on average. 

Some overhead past MSRP is normal, as partner models always include boards with quality of life features (like better acoustics or thermals) that may warrant a higher price within reason. These percentages are reasonable.

Looking at the data as a whole: The combined AVG % over MSRP for all in-stock GPUs collectively has decreased from 27.8% over MSRP in June to 10.7% currently.

In any case, broadly speaking, prices are getting better. There’s still a lot to be desired, but we're optimistic that things are at least moving in the right direction and we are pessimistic that this will soon stop.

GPU Availability Comparison | June to Current | GamersNexus

GPU Availability Comparison | June (6/12) to November (11/26) | GamersNexus

This table shows how the number of total and MSRP listings in stock for each current generation graphics card has changed from June to the present.

Total in-stock listing count increased by 26.7%, from 135 to 171 model GPUs since June, with in-stock MSRP listings up 346.2% to 58, from 13 previously. This is heavily contingent on when stock is checked, obviously, and we don’t know the total volume difference, but that will align with the total volume difference.

Intel’s Arc B580 and B570 both saw reductions to their in-stock listing total due to Weeliao’s removed Newegg listings, which we think is a good thing and was for the best. The 5090 also had a noticeable decrease in listings, down from 9 in June to only 4 at the time of data compilation.

GPUs that saw the greatest increases to in-stock listings include the RTX 5070 Ti (read our review) at 27 from 12, then the 9070 XT, and 16GB 5060 Ti and 9060 XT, each with 8 to 9 more in-stock models than previously.

The 5070 (read our review) and 8GB 5060 Ti each gained 10 in-stock MSRP cards, for the greatest improvements in this category. That’s good to see at least, although the 8GB 5060 Ti is probably because no one wants it.  

At the time of writing this, all current gen GPUs, excluding the 5090, are available for MSRP, up from 3 of 13 GPUs previously. We won’t adjust this after writing even if it changes, as it is important methodologically to lock-in the comparison at a fixed timestamp for every device.

Overall, GPU availability has definitely improved since June.

Moving on to the 9070 XT’s troubled pricing history:

9070 XT Price History

9070 XT Newegg Price History | March (3/7) to June (6/12) to November (11/26) | GamersNexus

This table compiles all 9070 XT listings we’ve pulled from Newegg since the GPU’s launch, with AVGs inclusive of all models and prices collected in March, June, and November of this year.

During its release, the 9070 XT’s AVG price was $729, or 21.5% over MSRP.

From launch to June, AVG price increased to $787, or 31.2% over MSRP. Things got worse, making AMD’s MSRP promise, in our eyes, a flat-out lie.

Since then, the 9070 XT’s AVG decreased, resulting in an all-time low AVG of $660, or 10% over MSRP currently. That’s down from 31% over in June.

The 9070 XT’s availability also considerably improved since this summer, growing from 3/15 to 12/14 listings in-stock.

Unfortunately, this isn’t a perfect comparison because the models listed on Newegg changed each time we collected the results, as indicated by gaps in our table. 

That’s why we made this one:

9070 XT Price History (Narrowed)

9070 XT Newegg Price History | March (3/7) to June (6/12) to Current (11/26) | GamersNexus

In this table, we’ve narrowed our results to only include the 8 specific 9070 XT models that remained listed throughout each of our data captures. Naturally, the results seen here largely correspond with those from the previous table.

Each listing’s price increased or didn’t change from launch to June and then decreased in the time from June to now, with AVGs of $720 initially, $801 this summer, and $675 currently.

Generally speaking, aside from the summer spike, this aligns with how pricing typically changes over time -- especially for AMD GPUs that tend to launch a little too close to NVIDIA pricing.

AVG price being currently lower than it was at launch is promising, but we only saw one 9070 XT model at MSRP at the time of data compilation.

We’d like to see more from AMD on this front, especially considering the company’s extremely defensive claims about it being inaccurate to call the 9070 XT’s MSRP a launch-only pricing. AMD is twisting reality with its prior statement and has now been objectively proven to be false, making them, in our opinions, either liars or manipulators, if not both. They would have done better to keep their mouths shut.

Now, moving on to some of our less conclusive findings:

Number of Listings per GPU vendor

This chart represents the total and in-stock listing count for each GPU vendor’s current series GPUs on Newegg.

Out of the 243 total current gen Newegg listings, 192, or 79%, were provided by NVIDIA. NVIDIA is claimed to have around 90-94% of GPU marketshare, so this somewhat aligns with that.

Of NVIDIA’s 192 listings, 133, or 72.7%, were in stock.

AMD partners provided 42, or 17.3%, of Newegg’s current GPU listings, with 39 out of its 42 in stock (meaning 93% in-stock).

Intel partners supplied nine GPU models, making up just under 4% of all current gen Newegg listings, with nearly 90% of its models in-stock.

Obviously, we’d like to see more competition from AMD and Intel here. Both companies have a lot of work to do if they ever hope to challenge NVIDIA’s dominance over the GPU market in any meaningful way. There’s a game here of flooding the shelves just to be the most present by volume, which helps with consumer perception but also just the odds of someone clicking on your model. It’s a chicken-or-egg problem for the non-NVIDIA, though, because without more market share, it’s hard to justify more model GPUs. 

“AVG Low Price” - MSRP

This chart explores a different approach for a metric that’s representative of around the cheapest price you could realistically find any of these new GPUs for. We wanted to eliminate overpriced partner models.

We came up with what we’re calling “AVG Low Price,” calculated using each GPU’s 4 cheapest, in-stock, first-party listings. This is particularly useful for excluding each GPU’s most overpriced models.

This chart illustrates the difference between each GPU’s “AVG Low Price” and its MSRP.

“AVG Low Prices” for the 5060 Ti 8GB, RTX 5070, RTX 5060, 9060 XT 8GB, and RX 9070 (read our review) were all below their MSRPs, ranging between $7.50 and $32.50 less.

Continuing down the chart, we noticed that the majority of current gen GPUs had “AVG low prices” basically equal to MSRP. This is a strong indicator that prices are gradually getting better, generally speaking. It’s just unfortunate as this is likely temporary due to the memory situation.

Finally, taking the cake as the most egregious example on the chart: The RTX 5090’s “AVG Low Price” was $1,100 over its $2,000 MSRP, again highlighting NVIDIA’s lack of competition in the high-end. 

Conclusion

Our conclusion today is simple: As a whole, the GPU market, from a pricing standpoint, looks better now than it did when we collected prices in June, at least for the time being.

In-stock AVG price has decreased for all but one current series GPU on Newegg.

Count of in-stock listings rose from 135 to 171, and those available for MSRP increased from 13 to 58.

Previously, the 5060, 8GB 5060 Ti, and 8GB 9060 XT were the only new GPUs available for MSRP on Newegg. At the time of writing this, all current series GPUs have MSRP listings in-stock except for the and 5090.

Unfortunately, the 5090 was largely unaffected by the market’s improvements and continues to stand out in terms of the most abysmal price hikes over MSRP.

We may check back in a few more months specifically to look for any meaningful price reductions to the 5090, but we generally feel like these are the most stable prices we’ll see in the GPU market before the DRAM price increases start getting passed along to the consumer.

We're finally benchmarking GPU performance in Linux, first using the Bazzite OS following thousands of community requests specifically for this operating system.

The Highlights

• A lot of our testing here is exploratory and research for establishing methodology, so we're still learning how to control the platforms and software for this benchmarking.
• Linux still isn't for everyone. Some users, like our own production machines, are bound to Windows by compatibility requirements with certain software.
• Gaming has dramatically improved on Linux over the years and is developing fast (despite still having issues), and so it may slowly start to become more viable for gaming users.

Table of Contents

• AutoTOC

Intro

Microsoft’s business model is built on spying and data acquisition, making the obstacle of learning a new operating system less towering than it used to seem. It’s time to start benchmarking in Linux to get data available for those jumping ship.

Editor's note: This was originally published on November 25, 2025 as a video. This content has been adapted to written format for this article and is unchanged from the original publication.

Credits

Test Lead, Host, Writing

Steve Burke

Testing, Writing

Patrick Lathan

Testing

Mike Gaglione

Camera, Video Editing

Vitalii Makhnovets

Writing, Web Editing

Jimmy Thang

With the enshittification of Windows and the rise of SteamOS, the requests for Linux testing are coming from more than just a few passionate Level1Techs viewers. 

It's been a lot of work, but we have some initial GPU results for Linux, specifically for the Bazzite distribution. The testing was about 3 weeks of nonstop work, plus about 2-3 weeks of bench setup and “bench hygiene” management. We’re still learning. Benchmarking anything to the level of isolation we maintain requires a lot of controls and experience, so you’ll have to work with us while we learn these processes. 

This first round of testing has mostly been for our own training, which is what we’ll present here. So first we'll cover some of the bumps we ran into as first-time-in-a-long-time Linux users. We spoke with some of the Bazzite maintainers during this as well, which was helpful for getting test controls in place.

Linux still isn’t for everyone, but it’s also not hard to throw it on another SSD and test drive it. In particular, users with applications that may not have Linux adaptations may find themselves stuck on Windows. This includes us for our main production machines, as we just use too much software that needs it. But for users who mostly game and use a web browser, it’s a good time to experiment with Linux.

During this process, we had crashes, freezes, issues with anti-cheat, and some insanely long shader compilation times, but in spite of all of this, Linux is working better than it has in its history for gaming. If more users gritting teeth through initial learning pains means getting away from Windows and its spyware, we count that as a win. Overall, Linux does work better than it used to, but it’s not perfect.

The disclaimer first is that our level of testing for GPUs and CPUs on Windows is extremely refined and rigid. We’ve also been highly automated for almost a decade now, including capturing images during test passes for image quality checks. All of that goes away with a new operating system, as even the test software is different. This means that even just the unit of measurement we put on a chart could feasibly be calculated slightly differently.

The point is that we’ve taken a lot of care to do this justice, but we want to be fully transparent that Linux is almost by design a non-rigid solution that gets regular, sometimes impactful updates. That makes testing hard. We’ll do our best, but work with us as we learn this process.

Test Setup

Methodology is important here since it’s all new and sensitive to change.

The first thing you need to know is that this is not directly cross-comparable with our Windows testing. This is for a lot of reasons, including changes to test methodology, but also measurement tools and the games themselves not necessarily rendering the same thing, the same way on both platforms.

With all testing, we have to strike a balance: we need to be thorough, but we also need to limit our scope so that we can finish testing and publish something before more major changes come down the pipe.

GPUs are the most straightforward hardware category that we test, so we began with our standard GPU test suite on our standard platform.

We knew we needed to start with a single distro. The frontrunners were SteamOS (which we already have experience with on the Deck), CachyOS, or Bazzite. SteamOS was our first choice, but it's still not officially available for desktop. CachyOS should have more cutting-edge performance than Bazzite and might be more flexible, but potentially at the cost of some stability. Bazzite is popular, it's built for gaming, thousands of you requested we test it specifically, its developers have made executive decisions that simplify our own choices and have contacted us, 

And, most of all, Wendell of Level1Techs recommended it as an easier-to-test distribution for us.

Testing is a little different from use. We need a test platform to be predictable, stable, and most importantly, we need it to minimize judgment calls for things like settings and tuning. If we start making exceptions for tuning on one device, we have to do it on all of them. Testing something closer to usable-out-of-box is going to help railroad us into a fair configuration for all vendors, as we can always fall back on their own rigid decision making.

Bazzite Desktop Edition comes in two main versions: 

NVIDIA and non-NVIDIA. 

After speaking with both the devs and Wendell of Level1 Techs, we set up one drive for each. The OS comes fully stocked with drivers and game launchers, so we didn't need to do any significant configuration after installation, which also means that our test benches reflect the default user experience.

Bazzite is an atomic bootable container image. There's a full explanation on the FAQ page, as well as advantages versus SteamOS, but the bottom line is that you can just grab an image and slap it on your system. Bazzite's goals are stability and compatibility, and recoverability in case those first two don't work out. It's intentionally not the highest-performance, most bleeding-edge OS, but it's low friction for beginners, and it should be less prone to "oh shit" moments where it turns out some feature was broken during testing and without our knowledge.

According to Antheas from the Bazzite dev team: "We used to include some gaming oriented kernel optimizations but, as of August, we do not anymore. We found that most of the testing for those changes involved running e.g. Rocket League at 600hz while doing something like compiling a kernel. Nobody does that or cares, and we found those optimizations could cause a 2-30% performance degradation, where the higher end was on Intel CPUs with heterogeneous cores.

He continued, “This actually ended up hurting our reputation which is very unfortunate because the version of one of the optimizations we used had a bad bug. Moving forward, we will no longer make these sorts of changes, so performance on Bazzite should mirror Fedora and Arch, but with more stability, as we skip minor versions and delay the major a bit."

During a test cycle for Windows, we pause Windows updates until all tests are complete. It typically takes us about 1 week to complete a full test cycle, so this isn’t a problem. Linux has a lot more layers that may update at a faster pace, plus our testing is slower since it’s less automated.

Our instinct is to freeze versions as much as possible without modding. Linux has a number of pieces of software that update regularly, including:

• The operating system itself
• The games, which get potentially different updates from Windows
• Steam, which gets meaningful updates
• The Proton translation layer
• The drivers
• Among others

The rapid progress of gaming on Linux complicates testing in a way that is more challenging than an actual user environment. Freezing some updates but not others could lead to issues where things break for testing, but not for a user.

Our compromise has been to freeze Bazzite (including drivers) and Proton during the test cycle and take detailed notes on everything else. In the future, we may shift to GoG, which allows using specific game versions. 

In general we've accepted that we'll have to throw away data and start from scratch more frequently on Linux.

We don’t talk much about the business side of things, but this testing sort of requires some insight there just so everyone understands how this complicates things.

For us we typically, if we didn’t do it the right way, lose money on a full, fresh-from-scratch test cycle for GPUs and CPUs just because we test so many. 

For instance, we spent 85 hours overhauling our GPU test suite in December of 2024. That included new games, benchmark validation, calibration, running about 26 GPUs through it, experimenting with new methods like early Animation Error metrics, and then working through to eliminate bad data and calculate things like standard deviation. If we were to then only publish a single video with that 85 hours of work, so that means adding more time for us to write, film, and edit it, it would not be possible for us to make money on that single project; however, if we then publish a follow-up GPU revisit or review that requires appending only one or two devices to the other 26, we can start recouping the costs of the groundwork.

For that reason, simply “rerunning” all the tests every time we want to add one device to the Linux suite is just not really viable without some major structural changes to our test processes. In other words, Windows makes it easier for us to hold data for longer because it doesn’t drift much without some major revision somewhere, but with Linux, we’ll need more frequent updates. Without a compromise where we can hold data for longer with Linux, we probably just need to do this a couple times per year at most with a brand new data set each time.

Anyway, that’s a lot of boring business stuff to say that it’s harder to find creative ways to subsidize a full suite of Linux tests than it is for Windows. We’re basically doing this because we think it’s important, we hate Microsoft and Windows, and we hope that investing in Linux testing early will result in easier processes later.

To cover the equivalent settings that we'd change on Windows, we set "Adaptive Sync" to Never, allowed screen tearing for fullscreen windows, and set the power profile to Performance. We've had warnings from Wendell and others about kernel spinlock performance loss, the CPU performance governor, and plenty of other Linux-y things, but the advantage of a plug-and-play OS like Bazzite is that we can just install it and run it. Any issues are a valid part of the experience.

There are a lot of ways to fuck this up. We think we’ve protected against them, but this is all new to us.

We're starting with our existing GPU suite for the most part, including the same settings we use on Windows, but we may change that over time. These results should only sparingly be cross-compared. The entire nature of this test is that it is not cross-comparable with Windows.

We're going to limit comparisons of Linux and Windows data. Linux framerate determined by MangoHud versus Windows framerate determined by PresentMon feels like an apples-to-oranges comparison, both because of the difference in tools and the difference in platforms. 

We're still getting used to how Linux handles frames; for example, we always allow screen tearing, but the concept of tearing is still new to Wayland.

The scope of our GPU test suite doesn't include extended stability testing. We'll report any crashes or incompatibilities that we encounter, but we aren't going to validate that Dragon's Dogma 2 can run for five straight hours on Linux without crashing, so be aware that there may be issues beyond what we report here. Our focus is on numbers today, not stability. 

We know a lot of you just want numbers, so we’re going to do that now with the caveat that there is a lot of critical methodological information that we’re showing after the charts. This includes changing game APIs and talking about native Windows vs. Linux versions of games.

All that we really ask is that you please do not cross-compare anything with our Windows data set beyond what we do here, as it’s just not that simple. You might as well be comparing different games to each other. The goal is how well things perform on Bazzite.

Let’s get into the numbers, then we’ll return to issues during testing.

For the GPUs on the charts that have various VRAM capacities, we used 8GBs for the RX 9060 XT (read our review) and 16 GB for the RTX 5060 Ti (read our review). The only reason for this discrepancy is simply due to GPU availability as we managed multiple other tests during all of this and we'll continue to look at adding more of these if we do another round of Linux testing. But for now, we'll start with these.

Bazzite Linux GPU Comparison Benchmarks

Dragon’s Dogma 2 - 1080p/Max

We’ll start with the Windows build of Dragon’s Dogma 2 at 1080p/Max settings. For this testing, we used KDE Plasma 6.4.4, KDE Framework 6.17.0, Qt 6.9.2, Wayland, and 10.0-2e Proton, except the 9070 XT (read our review) and 5070 Ti (read our review), which ran earlier and were both experimental 10.0-20250919.

The 5090 (read our review) leads the chart for 1080p at 176 FPS AVG. Just like with Windows, we’ll need to look to higher resolutions to ensure it’s not CPU-bound. The 9070 XT is disturbingly close for NVIDIA, up at 140 FPS AVG and leading the RTX 5080’s 134 FPS result by 3.9%. Lows are about the same between them. The 5070 Ti suffers somewhat in 0.1% lows. On inspection, it had one pass with a 23 FPS 0.1% low, one with 61, and two at 96-97 FPS. The GPU experienced more variance than typical.

The 9070 (read our review) is about tied with the 5070 Ti, but had superior lows. The 9070 also leads the 5070 (read our review) by 20%. The 9060 XT outdoes the 5060 (read our review), but the 5060 Ti 16GB outdoes both.

Dragon’s Dogma 2 - 1440p/Max

1440p had the 5090 at 154 FPS AVG, leading the 5080 (read our review) by 37%. The 9070 XT runs at about the same level of performance as the RTX 5080, which is another good position for AMD despite seeing a decrease in its relative rank from 1080p. The 5070 Ti trails this, with lows proportionally scaling.

Current newegg prices have the 5070 Ti at $730 to $750, with the 9070 XT at $600 to $650. In this situation, the 9070 XT is improved on the 5070 Ti in both performance and price.

As for the 9070, which is around $550 to $570 on Newegg as we write this, it’s outdoing the $500 RTX 5070 by 19%. 

The 5060 8GB falls off an absolute-fucking-cliff here and becomes unplayable.

Dragon’s Dogma 2 - 4K/Max

4K draws the numbers down. The RTX 5090 now has a 44% lead over the 5080 (as compared to 37% at 1440p), while the 5070 Ti is roughly tied with the 9070 XT. The 9070 is just behind the 9070 XT, which improves by only about 10% on the non-XT model. The 5070 ends up below the 9070, with the 9060 XT at the bottom of the chart. The RTX 5060 was not capable of running this test.

Black Myth: Wukong - 1080p

Black Myth: Wukong is next. In Windows, this game heavily slants in favor of NVIDIA with ray tracing enabled. It’s more balanced in rasterization, which is what we’ll look at first, but NVIDIA is still generally a strong performer here.

In Linux and at 1080p/High, the RTX 5090 runs at 115 FPS AVG, but has extremely disproportionate lows that slam its 1% and 0.1% averages through the floor. These aren’t anywhere close to the average, indicating a bad experience despite a high average framerate. We’ll look at that in a frame interval plot in a moment; unfortunately, animation error metrics don’t yet work in Linux software that we’re aware of, so we can’t evaluate that.

The 9070 XT has a lower average framerate, but offers a significantly better experience than NVIDIA’s flagship RTX 5090. The 5090 has an average that’s 9% ahead of the 9070 XT, yet the frametime pacing on the 9070 XT breaks away.

The 9070 repeats this: It’s only about a 5 FPS difference in average from the XT and still retains better interval pacing than the 5090.

The 5080 follows the 5090’s behavioral patterns, as does the 5070 Ti, 5070, 5060 Ti, and 5060 -- this is just an NVIDIA problem with this combination of game, driver, and operating system.

Even the RX 9060 XT is a better experience than the RTX 5090, as its lows are more proportional to the average.

For that matter, Intel’s B580 (read our review) is more consistent. NVIDIA has issues with its flagship title with this combination of software.

Black Myth: Wukong 1080p Frametimes

Here’s a plot of the frame-to-frame intervals, previously known as “frametimes.” We’ve been fine-tuning some of this language since our animation error white paper we released.

The frame-to-frame interval in milliseconds for the RTX 5090 illustrates frequent spikes to 18-20 ms, up from a typical result of 8 ms. That’s a 10-11 ms excursion frame-to-frame and it’s frequent, meaning that you can perceive a slight stutter. This is an instance where we’d love to have animation error metric in Linux, because it’d be a perfect way to drill-down into the issue.

For a 5090, this experience is unacceptable. We know it doesn’t do this in Windows, so this is unique to the situation.

The RX 9070 XT plots with extremely consistent frame-to-frame intervals, with almost no excursions. The few that exist are no more than 3-4 ms from frame n-1. So, while the 9070 XT has overall lower framerate or higher frame interval timing, the experience is far better when using it.

Black Myth: Wukong 1440p

At 1440p, the RTX 5090 establishes more distance in average framerate from the RX 9070 XT, yet still suffers from dismal 1% and 0.1% low averages. The end result is that the 9070 XT has remarkably consistent frame-to-frame pacing with some of the best proportionality we’ve seen on these charts so far, while the RTX 5090 Founders Edition flounders its way through the pacing. The 9070 XT is superior in this test. That may change if NVIDIA is able to fix whatever the problem is; however, a fix would likely reduce the average framerate by pacing the frames more evenly. 5090 lows here are also inconsistent, meaning that they vary widely run-to-run.

The other NVIDIA devices suffer a similar fate, like the 5080. The 9070 ends up better than the 5080 by experience and is functionally tied in average framerate. The lows on the 5060 are more even with its average, as we appear to be hitting some kind of floor for the driver. For its price, the B580 is doing well as compared to the 9060 XT and RTX 5060.

Black Myth: Wukong 1440p Frame Pacing

Here’s the frame pacing for the 1440p results.

The 5090 shows the same behavior as before, just with a higher baseline time required at closer to 10-11 ms. The excursions adjust accordingly, now at 19-22 ms spikes, or a deviation of about 10-12 ms frame-to-frame in some cases. The rapid succession of these spikes is what makes the experience bad.

The 9070 XT looks about the same as before: The average time is a little longer, but the fluidity is far better.

Black Myth: Wukong 4K

Here’s the 4K bar chart.

The 5090 handles 4K better than anything else, so we’ve finally found a situation where the game load crushes the other cards enough that even an inefficient and poorly running 5090 can still outmatch them. The lows come up a little from the prior chart only because the inconsistency run-to-run is so high for the 5090’s low performance, which is just the nature of poorly paced frame intervals in general.

The RX 9070 XT now falls technically behind the 5080 in average FPS, with its lows still better than the 5080 but worse than the 5090. That said, they’re more proportional to the 9070 XT’s average, so may still feel more fluid despite being overall lower framerate.

Resident Evil 4 (2023) - 1080p

Resident Evil 4 is up next. We had to prune more outlier test passes for this one than other games just due to run-to-run consistency issues.

At 1080p, the 5090 leads at 360 FPS AVG, about 27% over the 5080. We need to see higher resolutions for scaling away from CPU limitations.

The 5080 and 9070 XT are clearly unconstrained, though. The 5090 makes that obvious. The 9070 XT is just behind the 5080 in all 3 metrics, giving the 5080 a 7.8% advantage in average framerate. Broadly speaking, this isn’t looking as favorable for NVIDIA as in Windows environments. AMD seems disproportionately favored here as compared to typically. To be that close to the 5080 is not good for NVIDIA.

The 9060 XT is about 48% ahead of the B580 as a result. 

As an aside, the 0.1% low standard deviation is very high in this test, although the average framerate was overall remarkably consistent.

Resident Evil 4 (2023) - 1440p

At 1440p, the 5090 drops to 316 FPS AVG and now leads the 5080 by 51%. That’s a huge gain over the 27% previously and is mostly thanks to imposing more of a GPU bind.

The 5070 Ti and 9070 XT are about tied with each other, but the 5070 Ti has lower 0.1% low marks due to a lower performance pass. This is real data though and is representative of the experience, so we’ve included it.

The B580 and B570 (read our review) don’t perform well (again), falling below the RTX 5060 and landing at the bottom of the chart. The cards are cheaper than everything here, assuming MSRP, so that’s somewhat expected. In the very least, Intel’s frametime pacing performance is good here, with some of the most consistent frame-to-frame intervals on the chart.

Resident Evil 4 (2023) - 4K

At 4K, the RTX 5090 now clears a ridiculous 66% lead over the RTX 5080, seemingly continuing to gain as load increases. It’s able to brute force its way to the top of the charts. 

Standouts include the RTX 5060, which actually did OK for average framerate at 4K if not for the awful 1% and 0.1% low metrics, which call our attention to bad frame-to-frame interval pacing.

The B580 and B570 are at the bottom again, but to their credit, are both better than the NVIDIA RTX 5060 result. 

The middle of the chart sees the 5070 and 9070 about 10 FPS apart in averages.

Starfield - 1080p

Starfield is up next. No one plays this game, so we figured we’d multiply 0 by 0 and test it in Linux.

The 9070 XT did phenomenally here by comparison to the 5090. Actually, NVIDIA is just straight-up screwed in this game, with bad performance as a result of the drivers and game interaction. Obviously, a 5090 is just better than a 9070 XT in basically every objective metric; however, in actual use in this test, its performance is worse than an RX 9070 and 9070 XT alike.

The 5090 is just 19% ahead of the 5080 here, with the 5070 Ti about the same as the 5080 and 5070 not far behind. This is a clear indication that something is just broken in this chain of drivers and software.

The game was overall extremely consistent in its average and 1% lows.

AMD comes away looking good here and overall, has thus far held more of an advantage in Linux benchmarking as compared to its relative positioning in Windows.

Starfield - 1440p

In Starfield at 1440p, the same situation unfolds: The 9070 XT leads the charts, followed closely by the 9070 non-XT. These two cards remain as close as ever to each other. The 5090 is next, then the 5080. This time, the 5090 leads the 5080 by 24%. That’s still not normal, but better.

Broadly speaking, NVIDIA is getting wrecked here. It’s not even close. The NVIDIA GPUs still hierarchically line-up against each other by name, but do not land where expected by actual performance differences.

Starfield - 4K

At 4K, the situation improves slightly for the 5090, but not in a way that matters much: The 9070 XT is still in front, now at 60 FPS AVG and still with lows that scale as expected. The 5090 trails, basically tying the 9070 XT. The lows difference is within variance of the XT. The 9070 flanks the 5090 on the other side, so overall, the higher resolution has created a scenario where it’s clawing itself ahead, but still not by an amount you’d come to expect from its hardware.

This looks like how Intel GPUs sometimes feel on Windows: The hardware is clearly capable of more, but it can’t break out of some external box.

Speaking of Intel, it’s not present here because it’s just broken in this title. We’ll look at that more in a bit.

Dying Light 2 - 1080p

Dying Light 2 is next, first at 1080p.

The RTX 5090 leads everything here for average framerate, including about 43% lead over the RTX 5080. The lows for the 5090 and 5080 illustrate a limit to how close they can get to the average of NVIDIA’s higher-end devices, where we’re bouncing off of some kind of driver or software ceiling. The 9070 XT and 9070 both carry lows closer to the average, once again, with average results for the 9070 XT about tied with the 5070 Ti but better 1% and 0.1% low averages.

The B580 appears at 74 FPS AVG, showing that it’s playable and competitive with at least the 9060 XT.

Dying Light 2 - 1080p (5070 Ti) Frame Pacing

One thing here: The RTX 5070 Ti had extremely high run-to-run variance that appears to be just a part of the performance characteristics for this configuration, not some test outlier.

This plot shows that, overall, the frame-to-frame pacing of 4 passes is generally inconsistent. The spikes you’re seeing above each cluster of data are undesirable, but the worst one clearly happened just before frame 3000, where we had a spike to about 46 ms. This is long enough that you’d notice, even though it’s a single spike.

Dying Light 2 - 1440p

The 1440p results have the RTX 5090 50% ahead of the RTX 5080 in average framerate, with the 5080 leading a nearly-tied pairing of RX 9070 XT and RTX 5070 Ti. The 9070 XT maintains better frame interval pacing, although not by as much as some previous results.

The RX 9070 is likewise consistent. The B570 ran at 43 FPS AVG.

Dying Light 2 - 4K

At 4K, the RTX 5090 held a 119 FPS AVG and lows that were still disproportionately low as compared to those of the 5080. To us, this indicates again that there is a ceiling the higher performance NVIDIA card is bouncing off of for pacing.

The 9070 XT ran at 64 FPS AVG, roughly tying the 5070 Ti and landing within error of its 0.1% lows. Predictably, the RX 9060 XT, RTX 5060, B580, and B570 can’t really run this test.

Cyberpunk 2077 - 1080p

Cyberpunk 2077: Phantom Liberty is up now, tested first rasterized and without RT.

Immediately, the first thing we observed was a higher standard deviation in this test. The 0.1% low standard deviation was about 10-11 FPS, which is the highest of these charts so far. At least 3 cards were in this range. The game is just more variable than in our Windows environment.

The RTX 5090 ran at 195 FPS AVG, leading the 9070 XT by 32% in average framerate; however, the 9070 XT again has superior lows to the NVIDIA RTX 5090, and in a big way. Fortunately for NVIDIA, the card is still overall playable and large spikes aren’t easily noticeable. Because its lows are high enough overall, it may be the case that the 5090 maintains a better experience -- but only by brute force.

The 5080 trails the 9070 XT for average framerate and significantly in low averages. The 9070 is close in average FPS to the 5080 and advantaged in lows.

Intel’s B580 ran at about 60 FPS average, which isn’t bad for the Arc card. That has it behind the 5060, which leads by 20% in average framerate, but is roughly tied in 1% and 0.1% lows.

Cyberpunk 2077 - 1440p

At 1440p, the 5090 held a 148 FPS AVG with lows at 104 and 90 FPS. This is a 46-50% lead over the RX 9070 XT and RTX 5080, which have about the same average framerate. Lows are also better numerically, though less proportionally spaced.

The Intel B580 drops to 40 FPS due to the heavy load of 1440p, following the RTX 5060. The 5060 struggles with the combination of VRAM limitations and NVIDIA’s software-specific issues in this test, on this OS.

Cyberpunk 2077 - 4K

We don’t expect the RTX 5060 to be able to play this game at 4K, to be clear, but other than this result, the rest is similar to what we’ve seen previously. NVIDIA’s results are improving at the low end, but are still disproportionate in the 5090 as compared to other devices. The 9070 XT struggled at 4K, landing at 48 FPS AVG and trailing the 5080 in all 3 metrics.

The B580 and B570 both massively outperform the RTX 5060, as does the 9060 XT, particularly in low metrics. 

Baldur’s Gate - 1080p, Linux Native (Vulkan)

We’re getting into Baldur’s Gate 3 benchmarking now. This game is different from many of the others: It has a Linux native build and a Windows build, and its test matrix gets more complex by the presence of Vulkan and Dx11 support. The Linux native build had some serious issues that we’ll talk about below, but we still tested it.

At 1080p and with the Linux native build running the Vulkan API, NVIDIA runs completely sporadically and unreliably. We’re well aware that the 5090 is below the 5070, and that obviously this is not how these cards actually behave. After re-tests and evaluations though, the only thing consistent is inconsistency when it comes to NVIDIA with the Linux native build of Baldur’s Gate 3 with Vulkan. 

The 9070 XT ran at 238 FPS AVG here. Its lows weren’t proportional to that average and are overall less consistent than we’d like. The 9060 XT lands predictably below that, followed by a host of unreliable NVIDIA results.

Broadly speaking, this build is buggy. Due to issues with reliability, we’re not listing error bars on this one.

Baldur’s Gate - 1440p, Linux Native (Vulkan)

1440p is also messed up, as you’d expect. The results for NVIDIA are, again, unreliable and it is clearly just straight-up broken. The 9070 XT isn’t as consistent as it should be, though at least runs an average framerate that’s OK.

Let’s move on to a different build.

Baldur’s Gate 3 - 1080p, Windows (Dx11)

With the Windows build in Bazzite running Dx11, things look more normal: The RTX 5090, RTX 5080, and RTX 5070 are bottlenecked on CPU performance, with the 5060 Ti also bottlenecking but not as hard.

The 8GB RX 9060 XT ran at 130 FPS AVG and had lows similar to the RTX 5060 8GB card, both significantly ahead of the Intel B580. Even still, Intel manages 79 FPS average and overall OK lows when considering its generally disadvantaged position. The B570 isn’t far behind and matches the scaling we’d expect from prior Windows scaling comparisons of the two Intel models.

Baldur’s Gate 3 - 1440p, Windows (Dx11)

At 1440p, things are the same, just scaled down: The top 3 cards are hitting a CPU or other bottleneck, with the 5060 Ti occasionally bouncing off of it but still somewhat struggling on its own. 

Baldur’s Gate 3 - 4K, Windows (Dx11)

At 4K, the RTX 5090 hit 170 FPS AVG, which is technically ahead of its 1440p result. That’d make no sense in a GPU-bound scenario when going from 1440p to 4K, but in this case, the results at 1080p, 1440p, and 4K are all bottlenecked on non-GPU components for the RTX 5090, possibly also including software. The end result is that the results will be slightly more variable due to different constraints. Effectively, all 3 resolutions are identical for the 5090 because of external limits.

We were able to get the 9070 XT to run in this one. Its result was 103 FPS AVG, whereas we saw 115 FPS AVG with the Linux build and Vulkan, although not directly comparable. 

Ray Tracing Benchmarks

Next, we’re moving on to ray tracing benchmarks.

Cyberpunk 2077 - 1080p, Ray Tracing

NVIDIA has always had an advantage in Cyberpunk on Windows, and that sustains here -- a lot of that is just from the actual hardware itself.

The 5090 held a 135 FPS AVG, leading the 5080 by 46%. The 5070 Ti follows, with the 9070 XT showing up at 72 FPS AVG and landing between the 5070 and 5070 Ti. Intel’s B580 actually did well here with its 58 FPS AVG, planting it just ahead of the 5060 Ti. That’s a compelling argument for Intel.

The RTX 5060 fails completely and cannot run this test.

Cyberpunk 2077 - 1440p, Ray Tracing

1440p is predictably heavy for these cards. The 5080 is the last card to technically perform above 60 FPS AVG, with the 9070 XT approaching the 5070 Ti but not meeting it. In this situation, AMD does not have an advantage in low performance and is not more consistent in frame pacing than NVIDIA.

Intel, however, is disproportionately good for the value of its hardware and approaches the RTX 5070 while maintaining roughly equal or slightly better lows.

Cyberpunk 2077 - 4K, Ray Tracing

4K is too heavy for this configuration and these cards, but we still ran it. The 5090 maintains actually good frame interval pacing here, with its lows relatively close to its average. Still, it’s at 49 FPS AVG.

The next closest card is the 5080, and that’s only at 30 FPS. 

Black Myth: Wukong - 1080p, Ray Tracing

We’ll just do these last few RT charts.

In Black Myth with RT Medium and FSR Quality, the RTX 5090 ran at 113 FPS AVG, leading the 5080 by just 30% and roughly tying in lows. AMD’s first appearance isn’t until below the RTX 5060 and, even then, its low performance is disproportionately bad. In this situation, AMD is getting comparatively crushed.

Overall, this is something of a reversal from prior charts. AMD has disproportionately bad lows and overall performance, while NVIDIA outdoes it. The RTX 5060 struggles in average framerate, but did OK in its lows. It seems to float closer to whatever barrier exists in these scenarios.

Dying Light 2 - 1080p, RT

Dying Light 2 is the last one. In this one with ray tracing, the 5070 Ti could not complete the test due to constant freezing and crashing that we were unable to work around.

In this test, NVIDIA again outperforms AMD significantly. The RTX 5070 outdoes the RX 9070 XT here, with only the 5060 Ti and 5060 behind the 9070.

Bazzite Summary

Overall, most of the games we tested ran without major issues on high-end mainstream hardware, and the Linux gaming environment is changing so rapidly day to day that any individual bug we experienced could be gone tomorrow. Intel cards and cards with less VRAM (like the RTX 5060) tended to have more problems. 

Before running these tests, we were having a hard time determining whether people consider NVIDIA's Linux support bad in the sense that cards don't work, or bad in the sense that NVIDIA's approach doesn't align with the spirit of Linux. 

From Antheas again: "In the context of NVIDIA drivers, ‘open’ refers to the kernel module, which is now under the MIT license, but the drivers are still proprietary. Before the open module, there was a closed module and the main difference between them is that the closed module was made for pre-GSP [GPU System Processor] cards. Before GSP, the NVIDIA driver did a lot of heavy lifting and e.g. had 10k lines of Monitor EDID fixes.”

The Bazzite developer added, “Stuff like this is important proprietary information for them, so the driver was closed source, which made a lot of Linux users upset and was a legal gray area with the kernel code. Now, with GSP, which is a co-processor on the card, they can shove all those fixes in the firmware of that co-processor, which is closed source, and distribute an 'open' driver that communicates with that. It's just fancy software-engineer lawyering."

For gaming, the NVIDIA Bazzite build worked as well for us as the non-NVIDIA build, at least with the relatively new GPUs that we tested.

It was also just an awesome change to be contacted directly by the developers of Bazzite, MangoHud, and pretty much any open-source tool we mentioned publicly in the course of working on this. On Windows, the OS and the driver are out of our hands and Microsoft employees only ever contact us under strict anonymity requirements, but on Linux, we can directly ask for information and request features.

This alone is a great sign.

Keeping it real with everyone, it’s not perfect and you’ll need to set your expectations.

Testing Problems

We can thank the Steam Deck (read our review) for normalizing how well so many games run through Proton. Although we had issues in some instances, this is still largely true.

As Steam Deck users are aware, some anticheat software flat out refuses to run, which means games can refuse to run. We dropped F1 24 immediately for this reason. This is the only game from our suite that was affected, but since anti-cheat is usually associated with popular mainstream games, this remains one of the most significant obstacles to Linux gaming.

Our two RE Engine games, Resident Evil 4 and Dragon's Dogma 2, didn't show options for ray tracing. It's a known issue and may be possible to work around, but not without a level of tweaking that we want to avoid. We're skipping RT tests for these two games on Linux until Capcom figures it out.

A few Steam games have native Linux branches, which Steam always prioritizes for download on Linux. Unfortunately, sometimes the native Linux versions suck. In the case of Total War: Warhammer III, the Linux build lags significantly behind the Windows version. It looks like most users opt to run the Windows build through Proton instead, so that's what we tested.

Baldur's Gate 3 is another game in our suite with a native Linux version, although it only came out after we had already started testing, and promptly broke desktop Linux installs because Larian forgot about non-Steam-Deck Linux users. 

This distributor-side messiness will become more common as the Steam ecosystem expands. 

We also found that full-screen mode in Baldur's Gate 3 would always run at the native display resolution regardless of in-game settings (specifically with the Linux build).

With Cyberpunk 2077, with our RT Medium settings profile applied, we'd consistently get crashes on the second loading screen. We could load a save once, but trying to reload or exit to the menu would crash the game. With our non-RT settings profile applied, this didn't happen.

We had issues across multiple games that may have been caused by low VRAM. For example, the RTX 5060 was able to struggle through Total War: Warhammer III testing at 1080p, but completely failed to get past the benchmark loading screen at any higher resolution. 

Intel's cards had the most problems, although we were told by Bazzite's developers that Intel GPUs would be the most affected by updates, so pausing updates may have prevented some fixes, but that’s the tradeoff we had to make. Using the B580, the native Linux version of Baldur's Gate 3 wouldn't launch, Dragon's Dogma 2 couldn't run long enough to test, and Starfield couldn't load saves. We also saw flickering and artifacts at desktop. None of the games in our test suite showed RT as an option with Intel except Cyberpunk 2077, and when we moved down to the B570, it couldn't even run Cyberpunk with RT without crashing. To be fair, Cyberpunk with RT crashed consistently on the RX 9060 XT as well.

We also had multiple issues that may have been related to MangoHud. The most obvious was that games would frequently crash to desktop when MangoHud finished logging. We've been in contact with MangoHud's developer and have some ideas, but for now we're just dealing with it. We also had games freeze for up to a minute at a time, which could be worked around by tabbing out and back in. 

We intentionally hid the MangoHud overlay during all testing because of an earlier issue on the Steam Deck where any overlays would cause stuttering and other issues.

After a certain point, MangoHud stopped working with the Heroic Games Launcher (which we were using to run Cyberpunk 2077). This was alarming mainly because it implied that something was updating outside of our control, although it may have been due to freedesktop's policy of marking releases EOL after two years. Someone on Reddit had a related problem that pointed us to a fix, and then they deleted the post, so good luck.

Steam's first-time shader compilation step took an extremely long time, especially with NVIDIA cards, up to 25 minutes for Starfield on one of the NVIDIA cards we tested. Many of the Steam games we tested forced tiny shader cache updates every day, which is behavior we've seen on the Steam Deck as well.

We test nearly every single button in Intel's graphics software, track 641 bugs across 10 months of updates, and more

The Highlights

• We're revisiting Intel Arc drivers in 2025 to look at the post-Battlemage, pre-Celestial GPU driver situation
• We experienced one catastrophic crash that seemed to break the drivers
• Intel’s drivers have improved significantly, but we still wouldn’t recommend the company’s Arc GPUs to those not willing to do any troubleshooting

Table of Contents

• AutoTOC

Intro

It’s time to revisit Intel’s Arc GPU driver situation.

Editor's note: This was originally published on October 31, 2025 as a video. This content has been adapted to written format for this article and is unchanged from the original publication.

Credits

Test Lead, Host, Writing

Steve Burke

Writing, Testing

Jeremy Clayton

Video Editing, Camera

Tim Phetdara

Writing, Web Editing

Jimmy Thang

We built a gigantic spreadsheet to track 641 issues across Intel’s Arc GPU driver software for the past 10 months, compiled the driver release timeline to see how committed Intel is to new game releases and their speed of driver launches for those games, checked the game-ready driver launch lag, and clicked nearly every single button in Intel’s graphics software to see what features are broken -- and out of all of that, we only had one catastrophic crash that seemed to break the drivers.

And to be fair to Intel, that is a massive improvement from when we first tested their driver software.

Intel Arc GPUs have been around for over three years now, from the “Alchemist” A380 (watch our review) in July of 2022, to the B580 (read our review) and B570 (read our review) 10 months ago, to now. It’s been a real rollercoaster of ups and downs.

In this story, we’re focusing fully on Intel’s driver updates, patch cadence, and bugginess. This is our second revisit and third in the series to see how Arc is progressing. Instead of focusing on performance this time, we’re looking at how actually usable the drivers are since Intel is the newcomer to this space.

Overview

We have two sections to this: Intel’s driver update timeline and its bugfix consistency for one, then positive features, bugs, issues, and usability annoyances in Intel’s current graphics software.

Outstanding Issues From Our Last Revisit

Now we’re escaping the patch notes to revisit the issues that we specifically mentioned as unfixed or semi-fixed in our One Year Later piece. If any of these are still around after two years, we have a big problem. 

We’ll start with a recap of existing and prior broken features, bugs, and shortcomings to see where Intel has improved, then move to their driver update cadence.

Crash that Broke the Driver

Back when Arc launched, we experienced a lot of driver crashes and occasional driver corruption. That’s improved, but isn’t fully resolved.

This time, we had one really bad situation where IGS locked up and eventually crashed when we used the embedded link to the Windows display settings menu from the IGS Display tab. After restarting IGS, the performance metrics were humorously broken, showing extremely high numbers of “Raw” type, which you can't even pick under normal circumstances.

After rebooting, we didn’t get any video out to the monitor beyond the Windows 11 loading spinning graphic. This is when the graphics drivers are loaded, so it seemed the crash cooked the driver to the point of being nonfunctional. Reinstalling it fixed this, but we couldn’t replicate the problem a second time.

Happening once is still a problem. Intel has to fix these high severity issues of crashing and driver corruption long-term to win over users out of the enthusiast base.

It’s easy to say “it only happened once,” which is true, and a lot of our audience can deal with that. But this is the distinction between Intel and more established companies in graphics -- not that NVIDIA’s drivers have been particularly good this year.

Smooth Sync

Smooth Sync is up next, which is different from other forms of frame synchronization in that it doesn’t prevent tearing, but instead adds a dithering effect on the border between torn frames on screen. It originally had issues with visual corruption in several games, one of which was Rainbow Six: Siege. We showed this in our original driver check-in, leading to Intel disabling it in problematic titles. This was their “fix.” It’s certainly hard to have feature problems if the feature no longer exists in the game.

We decided to test Smooth Sync in Siege X.

To Intel’s credit, Smooth Sync works correctly in this title. The dithering also seems cleaner now than when we last looked at it. We aren’t convinced that it actually makes it nicer to look at, but what matters is it isn’t just broken. This issue seems to be resolved.

Speed Sync

Now for Speed Sync, which allows the game engine to run uncapped, but only the most recently completed frames are sent to the monitor when the monitor is ready. The theory was that this allows the latency benefits that come from uncapped frame rates, just without creating tears. NVIDIA’s Fast Sync is similar in concept.

Two years ago during our revisit, we found Speed Sync to be gone, but it’s back now.

Trying to use Speed Sync with Siege caused the game to consistently crash on launch. This isn’t mentioned as a known issue that we could find on the Arc patch notes, but there are sporadic reports of similar issues around the internet. Intel still has problems here.

Overlapping Apps and Settings

The software is up now.

Intel used to have both “Arc Control” in the driver suite and “Graphics Command Center” in the Microsoft store as a free application, both separate but sometimes overlapping in features and it generally was a completely awful experience because you had to go to the Microsoft Store to get some of the features.

In the time since, Intel cut it down to just “Intel Graphics Software,” which Intel migrated to at the end of last year, in line with the B-Series launch. This was a good consolidation to get away from the confusing mix of applications before.

VRR and DRR

One of the annoyances previously was that there were multiple locations to control Variable Refresh Rate. And there still are. 

The Display tab of Intel Graphics Software has a toggle for Variable Refresh Rate, the bottom of the Display tab has “Variable Refresh Rate Mode,” and then separately, Windows has its own “Dynamic refresh rate” under the Advanced display menu, which isn’t supported in this instance even though the same page says that variable refresh rate is supported.

Each of these technically serves a different purpose: The Graphics Software setting is only available when a GPU is connected to a VRR-capable display. The Variable Refresh Rate Mode setting is always available and controls whether VRR is forced on windowed applications in addition to the usual fullscreen ones. We think this option is buried and should be moved up to be alongside the first option. 

As for dynamic refresh rate, that’s distinct from variable refresh rate and exists to switch the refresh rate of the monitor globally. This requires VRR up to at least 120 Hz.

Driver Support Assistant Forced Install

Intel originally required users to install its “Driver Support Assistant” when installing the Arc Control software, despite having a checkbox that implied the ability to remove it. It was a forced install with the software. Around the time we reviewed the B580, we mentioned that Intel had fixed this.

DSA sits there to notify the user of new driver updates, and runs the risk of downloading them automatically.

This was a problem and annoyance for us on test benches, so we had to remove it each time.

Thankfully, it’s no longer a forced install and was fixed a little while ago. You’re able to freely choose any of the components and software that sits on top of the base driver. You still have to manually opt out of Intel’s data collection program, but at least you actually can now while still getting the software.

HDCP

HDCP is next: We previously had problems with being able to use a capture card while HDCP was on. We had to use a janky splitter in order to make it happen.

This time, we didn’t have that issue. We didn’t try a bunch of different combinations, but when we hooked up an Elgato 4K X capture card, HDCP support showed as “off” in IGS, and we were able to capture without any friction. We’d still like a manual toggle to control HDCP.

Notification Overload

At launch, Arc Control would blast the desktop with toast notifications, and it wasn’t possible to clear them manually. We just had to wait for them to disappear. During our last revisit, the number of pop-up notifications was reduced, but not eliminated.

Now with IGS, they’re almost totally gone. We caught only 1-2 pop-ups through all of the testing, down from dozens in our first round and a handful in our second. That’s a good move.

Cut Features

Over time, Intel has dropped a couple of features from its graphics software. We had specific criticism in our original review about how Intel was trying to do too much all at once, attempting to duplicate all the features of its competitors like it was knocking out a marketing checklist just for the sake of putting it on the box. They’re narrowing the scope now, which is good.

Intel no longer bundles recording and streaming functionality into the software. That’s fine with us – there’s no need for Intel to compete with OBS at this point when it has to deal with maintaining the software to ensure that it isn’t messed up and broken.

The previous overlays of any kind have been removed. No FPS counter, temperature monitor, nothing. That responsibility now sits entirely on PresentMon, which is its own utility. 

IGS has a tab for metrics, but it’s rough – more on that in a minute.

We previously included overclocking among the things that came off as unnecessary fluff at the time that Arc originally launched, especially since basics like the voltage setting were configured to the incorrect decimal point for the unit of measurement. Performance tuning is still in and they’ve fixed most of the basics – it also makes a lot more sense to us now versus during Intel’s first-wave efforts with the A-Series when all Intel really needed to get right was stability and game support.

Overall, we think Intel’s graphics software has improved from the simplification and consolidation of software.

Intel Graphics Software’s Features and Issues

As for the current state of Intel’s software, it’s broken down into multiple tabs: Home, Profiles, Graphics, Display, Performance, Video, Notifications, and Settings. We went through all of them, adjusting values, flipping toggles, and generally poking it to see if we could get anything to fall out in a general shake test.

In general, most things appeared to work correctly. For example, the video adjustments panel is so extensive that you can make Cyberpunk look like the image above if you really want to. 

Broken Performance Metrics

This next one is weird. IGS has performance metrics monitoring built-in, but it’s unreliable, often incorrect, and has very limited controls.

First off, it’s pretty clear that it uses an internal version of PresentMon for things like frametimes and latency, but we couldn’t find a way to point it at a target application. This should mean that it targets the app with the highest GPU load (like a game).

However, it failed to do so in our testing. Sometimes it showed nothing at all, and sometimes it clearly hooked on some unrelated process, showing frametimes that align with something running well below 60 FPS.

There are options to control the sampling window size (averaging period) and sampling period (the time between polling), but changing them affects the numbers being shown in oddly proportionate ways that don’t make mathematical sense.

For example, stock values of 1,000ms window size and 100ms sampling period resulted in “GPU Time” reading around 100ms average. Dropping the values as low as possible resulted in “GPU Time” of about 50ms average. Setting them all the way up to 5,000ms window size and 250ms sampling period resulted in about 250ms for average GPU time. If you picked up on that pattern, you see that it ends up reading awfully close to whatever the sampling period is set to.

There was even more. Setting the window size to a shorter timespan than the sampling period basically prevents it from working.

We tried it out on one of our test systems and ran PresentMon at the same time to compare the readouts. Sometimes, IGS failed to show anything. The rest of the time, it would be sporadic and clearly wrong, like in the image above, where it shows about 9 FPS AVG, when F1 24 was actually running at 170-ish FPS. 

Anyway, the currently useless state of IGS’ monitoring tab needs to be addressed. Intel should either get rid of it or make it work. The good news is that this isn’t critical to actually using the GPU day-to-day.

Video Capture

We briefly tested video capture with OBS on the B580. It’s capable of doing 4K at either of OBS’ non-lossless simple presets, but only if the GPU has headroom. If it doesn’t, the encoder instantly gets overwhelmed and performance falls off a cliff, similar to what you’d see in CPU encoding when the CPU is preoccupied. 

The card becomes overloaded, so there’s some forgiveness there, but NVIDIA has an advantage here.

Its NVENC solution has a more graceful way of handling capture concurrently with heavily loaded scenarios. Intel doesn’t seem to live up to that just yet.

Smart VSync

Intel’s inconsistency comes back into play with Smart VSync, which is intended to only turn on when the game’s frame rate exceeds the monitor’s refresh rate. It worked reliably when changing the setting prior to launching the game, but only worked one single time when we changed it while the game was still running; after that, it no longer functioned in this order or way, contributing to the overall confusion caused by Intel’s inconsistent driver behavior.

UI - Minor Bugs and Annoyances

We’re moving on to smaller complaints now.

While we were in the Graphics tab, we found a small UI bug under the Shared GPU Memory Override setting. The default setting on the machine we got this capture from was a memory limit of 57%, or 36.1GB. When we moved the slider, turned off the override, then back on, the slider was still where we had set it, but the value on the left had reset to 36.1GB. One of them is wrong. This is a bug.

Less a bug and more just annoying was the fact that after manually entering a value, the only way to lock it in was to click on the slider. Pressing enter, escape, or clicking away didn’t work. This was just a usability shortcoming. 

Also annoying was the UI for rearranging the order of the metrics on the Performance tab. You’re supposed to click and drag the item in the list, but it has extremely aggressive edge scrolling, making it difficult to place them where you want.

To wrap up our UI complaints, the VRAM Tuning tooltip doesn’t agree with the slider. The tooltip and the slider represent memory frequency with both GT/s and Gbps, which can be confusing for some users.

Frame Cap Limitations

The next one isn’t a bug at all, but is a limitation. Even though IGS has the ability to limit frame rates, it’s strangely limited to just DX9 and DX11 APIs – no DX10, DX12, or Vulkan. We’d assume that there’s some technical reason for the limitation, but from a consumer standpoint, it’s a shortcoming that’s unexplained.

It’s very easy to set global or per-game frame rate caps with NVIDIA and AMD, and they don’t seem to care about what API is under the hood.

Game Ready

Optimized drivers coinciding with new game launches are critical to the usability of any GPU for gaming.

NVIDIA has made a concerted effort to market its “Game-Ready Drivers,” which is one of its more meaningful campaigns and something we can get behind. NVIDIA’s strength is in this specific category, which is having a driver ready for the day a game releases. This requires direct cooperation with the developers to get early access and sometimes provide engineering resources for shader optimization, something we talked with Tom Petersen (now at Intel) about last year.

This is one of the more important areas where Intel needs to deliver.

In the time since the B-Series launch to when we stopped gathering data, Intel has included “game ready” support for 39 total games, 34 of which were mentioned as supported on or before the game’s release date. That’s not bad -- 87% of the total named games were supported in time for a player to buy it on launch day and have explicit tuning from Intel. 

Of the 5 that got mentioned post-launch in patch notes, the average is just 2.4 days of lag time. This is a significant improvement from Alchemist.

That’s a good ratio of “day zero” games, and pretty reasonable timelines for the few that aren’t. Ironically, the one with the longest wait was Silksong.

We aren’t aware of what Intel’s selection or submission process is for which games get an explicit mention like this, or what level of validation takes place, but based on the data, it seems like Intel is staying on top of new releases. They don’t get all of them -- Starfield at launch was functionally broken on Intel. But to be fair, Starfield at launch was also just functionally broken.

Driver Timeline

We’ll move to a timeline of the driver launches since Battlemage to get an idea for how committed Intel is to Arc.

We gathered archival driver release data from TechPowerUp and Intel, starting at the launch of the B-Series (version 6325, December 13, 2024) until roughly mid-September (version 7029), just before Intel made the main driver package Arc-only and kicked out the 14th Gen IGPs and older. It’s possible we missed a few smaller or less documented hotfixes in this.

Over the last 10 months from what we found, Intel has released 13 validated WHQL drivers and 16 beta drivers. Intel has some areas it pushed, like late May through June with 3 beta releases and 1 WHQL release.

On May 26th, Intel’s 6795 Beta introduced new game support and performance improvements for Elden Ring Nightreign, which would launch 5 days later, Dune: Awakening to prepare for its June 10th launch, and Monster Hunter Wilds, which came out in February. For at least this driver, Intel illustrated that it’s trying to have drivers ready in time to ensure compatibility with new games or improve their performance.

The 6874 WHQL release from June 2nd had Farming Simulator flickering texture fixes, so these issues still exist for Intel. It also had important power management updates for Lunar Lake Core Ultra 200V series CPUs.

Just two days later, Intel pushed the 6876 beta. It had game support for FBC: Firebreak, Rematch, Stellar Blade, and Rainbow Six Siege X. It also featured what is possibly the most vague fix we’ve ever seen. For the Arc B-Series GPUs, Intel said it fixed “Intermittent visual artifacts may appear in certain usage scenarios.” Ah yes, we fixed the problem that might happen sometimes in a subset of nondescript circumstances. Thanks, Intel.

2 days after that, Intel pushed the 6877 beta driver to address one specific issue tied to audio on Core Ultra Series 2-equipped handhelds. 6 days later on the 12th, they pushed a full release that only noted a fix for crashes in Overwatch 2.

The rapid, quick changes followed by a major driver version shows Intel is at least keeping busy chasing down individual bugs and fixing them as fast as possible. That’s a good sign of responsiveness. We think this is a great philosophy in terms of support that caters to the relatively small group of enthusiasts who have taken the plunge on Arc. It also shows that Intel is alive. So, it’s kind of a canary in the coal mine in terms of if the GPU support will stop.

We’ll quickly run through some light comparisons to NVIDIA and AMD.

In that same time frame, NVIDIA sits in some contrast to Intel’s approach, having pushed 21 total drivers (18 standard, 1 beta, 2 hotfix). The vast majority of the releases are standard WHQL, which makes sense to us considering NVIDIA’s dominant legacy position. Its drivers are mature and rarely require quick fixes -- except the beginning of this year, when they released hotfixes to fix hotfixes.

AMD’s Radeon driver team was quiet in comparison to both, having released only 11 drivers (5 standard and 6 beta). To be clear, the quantity of drivers isn’t an indication of the quality of the drivers, but considering the frequency of the game releases, the quantity does matter.

XeSS Support

XeSS support is next.

For better or worse, the industry’s favorite buzzword comes into play: AI.

Upscaling is actually an important feature at this point. AMD and NVIDIA have FSR and DLSS. We checked on Intel’s progress with XeSS adoption. 

This chart shows how many games support each given form of upscaling. Support for Intel’s XeSS isn’t as extensive as its competitors, totaling 334 titles. This data is pre-XeSS3 announcement, as that’s not fully rolled-out yet.

AMD nearly doubles Intel’s XeSS count at 624 games supporting some form of FSR. FSR 1 is significantly lower quality than the versions that succeeded it due in major part to being strictly spatial and not temporal, but we’re counting it here.

NVIDIA holds a numerical victory against AMD, with 649 titles supporting DLSS Super Resolution. 

NVIDIA’s introduction of DLSS overrides means that most DLSS titles should be able to be forced to run the latest models. Of course, the game matters more than just the number of games, but NVIDIA gets credit for having both bases covered.

XeSS 2 is an especially bad situation, with only 44 games officially supported on the list at the time of writing for a solution which launched alongside Battlemage in December. It’s starting to look kind of abandoned. “XeSS 2” is actually an expansion of the feature set to include frame generation and XeLL (low latency), and not a replacement for the actual upscaling in base XeSS. The additional 290 games we have listed under XeSS 1 count just as much when considering upscaling alone.

Also notable is that Intel recently opened up the possibility for NVIDIA and AMD GPUs to officially support XeSS Super Resolution and Frame Generation, aka XeSS-SR-FG, which is nearly a whole alphabet. Intel did this by publishing its XeSS 2.1 SDK. The only requirement is support for Shader Model 6.4, which applies to NVIDIA’s 3000 series onward and AMD’s 6000 series onward. In the time since, Intel has announced XeSS3, which we detailed in our Xe3 architecture article, which talks about their version of MFG.

Lossless Scaling

Of course, based strictly on numbers, Lossless Scaling’s LS1 wipes the floor with all of them, supporting a near-infinite number of games (if that’s the metric we’re judging by). That’s the benefit of not needing game or engine integration, though that does mean image quality suffers sometimes significantly. Although on a technicality, FSR1 can also have an infinite bar because Lossless Scaling can run FSR1, which is sort of like, "we heard you like upscaling so we put upscaling in your upscaling so that you can upscale while you upscale."

We have a whole deep dive on that you should go check out if you haven’t seen it.

Fixed and Known Issues

This table is deceptively small, considering that it required crawling through and collating 641 mentions of known issues and 81 mentions of fixed issues across 9 months worth of patch notes. 

This took a while to put together.

Intel declares 81 total issues fixed since the B-series launch – 12 for A-Series, 34 for B-Series, 4 for Core Ultra 1 integrated GPUs, 25 for Core Ultra 2 integrated GPUs, and 6 for software or general fixes. Only 44 of those total fixes were mentioned previously as known issues. Of those, the average time to fix was 39.9 days. We assume that the rest of the previously-unmentioned fixes were things that got noticed and addressed quickly, such as between releases.

We also noticed two issues that were mentioned as “fixed” twice for the same exact issue and for the same product category.

The first was first mentioned as a known issue on April 3, 2025 for the Arc B-Series GPUs, “Returnal (DX12) may experience an application crash during gameplay with Ray-Tracing settings turned on.” This was marked as a fixed issue on June 30th, but was then fixed a second time on July 23rd, “Returnal (DX12) may experience an intermittent application crash during gameplay with Ray-Tracing settings turned on.” The one difference is the word “intermittent,” leading us to think edge cases that caused Returnal to crash continued to surface.

The second issue was on August 7, 2025 for the Arc A-Series GPUs, “Mafia: The Old Country (DX12) may exhibit corruptions on character hair during gameplay.” It got marked fixed on August 11th, resurfaced, and got fixed again on September 10th. That second update narrowed the scope down to “certain scenes,” which sounds like the same kind of situation – Intel thought it fixed the problem, but somehow "corruptions on character hair” returned. Overall though, the important thing is the average time to fix an issue that’s listed as known, which was just under 40 days.

Conclusion

To wrap this up, we think that the usability and general annoyance level of Intel’s Arc GPUs are in a significantly better place than they were a year ago, definitely better than two years ago, and night-and-day from the original Alchemist launch. So they are gradually improving.

The fact that there’s only one piece of software (Intel Graphics Software) for all settings is a substantial improvement, and the fact that it’s a less bloated app than what came before is also good.

There are still things for Intel to fix, like the seemingly totally broken built-in monitoring and the unreliable frame sync settings. It can be frustrating to have game crashes just from picking Speed Sync, or from seeing Smart VSync work once, but never again.

There’s also things for Intel to upgrade, like adding frame rate caps for more than just DX9 and DX11 games, and if possible, making it easier to capture video while the GPU is fully-loaded.

We think that the more aggressive Intel stays on pushing fixes, the better off Arc will be. Overall, Intel’s drivers have improved significantly. We still wouldn’t recommend the company’s Arc GPUs to someone who doesn’t have a lot of patience or knowledge on how to do any troubleshooting, however. We don’t think the cards are there just yet. For most people in our audience, though, we think Arc is viable and worth considering if you’re able and willing to address potential technical issues if they arise. 

In terms of performance per dollar, the cards are in a pretty good spot.

Intel announced its Panther Lake mobile solution with Xe3 GPU architecture updates, building upon Xe2 and moving toward an eventual Celestial launch in the desktop GPU space

The Highlights

• Xe3 is distinct from Celestial
• We suspect that Xe3 sets the foundation for some of what Celestial will eventually get
• Intel's Panther Lake updates also include claimed power management and balancing improvements for CPU vs GPU power budget allocation in gaming scenarios

Table of Contents

• AutoTOC

Intro

Intel today is announcing its Xe3 GPU architecture to follow-up its Xe2 architecture. It’s not Celestial yet, but it’s getting close.

Editor's note: This was originally published on October 9, 2025 as a video. This content has been adapted to written format for this article and is unchanged from the original publication.

Credits

Test Lead, Host, Writing

Steve Burke

Video Editing

Mike Gaglione

Writing

Jeremy Clayton

Writing, Web Editing

Jimmy Thang

The biggest change to Xe3 is that it’s just larger, with render slices scaling up to more Xe cores per slice, an increase in L1 cache from 192KB to 256KB, a significant increase in L2 cache, and more registers that are better utilized. 

Micro benchmarks show significant improvements in occluded primitives culling for unnecessary triangles when rendering game scenes in addition to improvements in anisotropic filtering. 

Its variable register allocation and register changes also aim to unclog the pipeline so that the hardware can be better utilized, as one of the biggest problems with Arc in its current Xe2 and Battlemage implementation has been that there’s plenty of hardware, but it’s not getting used properly. This is a mix of both hardware issues, like with fixed function units in the architecture, and driver issues, which it has been slowly addressing. Some of this included moving off of emulation of things like execute indirect previously to eliminate overhead. 

For Xe3, Intel noted to us some of its driver improvements and software control panel focus as well, all of which should benefit the company as it moves toward its eventual dGPU Celestial GPUs.

This accompanies a number of other announcements related to its Panther Lake mobile solutions and laptop hardware, plus some “AI” and NPU hardware. 

We’re mostly going to focus on the IP block of Xe3 and the architecture and won’t be as focused on the product side for laptops. 

Although this isn’t a dGPU part, it’s likely that this approach will either be directly found in the next dGPU or will at least indicate which direction Intel is going.

Intel was clear that this isn’t exactly Celestial, which is the architecture following in the Alchemist - Battlemage - Celestial - Druid lineup. Intel noted that “Xe3P” will follow Xe3. The “P” unironically stands for “Plus,” showing old Intel habits die hard. Intel didn’t confirm this, but the impression we got is that Xe3P will be the “real” Celestial GPUs, while this Xe3 makes major changes that likely set the stage for it.

Overview of Announcements

Intel had a lot of announcements to share with the press for today. For our coverage, we’re focusing almost entirely on the Xe3 changes and micro benchmarks. We’ll cover some of the other news as well, like performance/Watt improvements and XeSS changes, but we’re not going to get into the NPU and AI processing changes today. There’s enough to talk about just with the stuff that’ll affect consumer desktop components in the future (plus the immediate impact to laptops).

All of this follows the announcement that NVIDIA is investing in Intel to build its own mobile parts with them later, but there’s no news on that topic today. This is all Intel’s hardware.

Naming Confusion

Briefly on the naming: Intel admitted its naming mix of Xe for IP and Alchemist / Battlemage / Celestial / Druid for branding has been confusing. It was careful to note that these parts are not Celestial and the impression we got was that they don’t want to burn the name on an incremental improvement prior to a pending major overhaul. Intel is sticking with “Arc B-Series” for the Panther Lake mobile parts, but is moving to the Xe3 architecture. Xe3P will likely be Celestial or desktop parts later.

Xe3 IP GPU Block

Intel specifically mentioned designing Xe3 to scale to larger configuration sizes, which would be good news for anyone who wants to see something higher-end than a B580-class card in the future.

Let’s get into micro benchmarks first, then look at the block diagram. 

This is a chart of micro benchmarks, which are workloads designed to target extremely specific functions or behaviors on a product. A 2x improvement here won’t equal a 2x improvement in most real-world applications, but these allow us to see where the improvements are appearing. Intel published these for Xe2 also.

In Xe3 for “depth writes,” Intel says it saw a 7.4x relative performance improvement normalized to clock frequency. We’re not certain, but our understanding is that this is not isolated for configuration size. This means that this isn’t a perfect comparison since the Xe core count is different between Xe2 and Xe3 in these tests. This 7.4x improvement outstrips the change in configuration size, though.

We asked Intel what “depth writes” means. The company told us that it’s related to high-Z culling and that this bar represents better primitives culling in the pipeline, meaning culling of unseen triangles and geometry sooner in the pipeline so as not to waste resources rendering unseen objects in-game. An example might be if a building is obstructing a player -- there’s no point rendering the player if it can’t be seen. Culling isn’t new and batching primitives in ways that eliminate occluded primitives has been around forever, but this shows that there’s still plenty of ground to gain here for Intel. This will result in better utilization of resources and allocating them to more productive work. Intel told us that the improvement to this process is disproportionately beneficial, meaning that it should have an impact in gaming performance that would be more noticeable than other improvements. We’d expect this to carry over to future Celestial dGPU parts as well.

The “High Register Pressure Shader” section also saw a large uplift in micro benchmarks at 1.9x to 3.1x. Scattered reads improved by 2.7x on the relative scale of time, with Intel noting to us that this has to do with using samplers to read data scattered across something like a texture (as opposed to a well-organized data set).

Mesh rendering is also shown here, with Intel telling us that Xe2 had already provided a proof of concept around improving mesh shading. Intel noted that this micro benchmark is representative of workloads where a lot of polygons are present, telling us that the uplift comes from a larger cache and more efficient use of its registers. Culling also contributes.

Quickly, Intel also saw uplift in anisotropic filtering, which is the old function that helps improve smoothness of textures and objects proportionate to the view frustum’s angle. Ray-Triangle intersection also improved by 2x in the microbenchmarks on the relative scale, which is noteworthy since Xe2 already benefitted from relatively large ray tracing improvements.

Looking back at the Xe2 micro benchmarks, Intel then highlighted Draw XI and Compute Dispatch XI primarily. At the time it talked to us about this chart, Intel told us that this was due to implementation of native execute indirect support for indirect draw and dispatch, as opposed to its Xe1 emulation of these functions.

Block Diagram

Time to get into block diagrams for how the new Panther Lake Xe3 block is constructed. This shows a 12 Xe-core configuration as the maximum size announced for mobile, with this configuration carrying 16MB of L2 cache, 2 geometry pipelines, 12 samplers, and 4 pixel backends. The L2 cache is noteworthy here.

This is the new Xe3 render slice. A render slice is Intel’s terminology that defines a block on the GPU containing Xe cores. For reference, the Battlemage B580 with Xe2 has 20 Xe cores on 5 render slices, so each slice is just one part of the total GPU.

The Xe2 slice had 4 Xe cores each, with Xe3 moving to 6 Xe cores per render slice. Intel also intends to scale-up the configuration size on mobile devices to a maximum of 12 Xe cores (or 2x render slices, up from 8 Xe cores on a prior 2-slice configuration).

The Xe3 render slice shows that each Xe core has 8 vector engines, which is unchanged from Xe2 cores; however, Intel is increasing the cache size in Xe3. Intel’s Tom Petersen stated, “The first thing we’ve done is increase the size of our L2. By increasing the size of the L2 from 8MB to 16MB, we reduced the traffic that hits the memory interface. That’s important because the memory interface is typically one of the most precious resources on a graphics chip. We can see anywhere between 17% and 36% traffic reduction heading towards memory, which has a significant performance effect on these different applications.”

Looking at Intel’s first-party results, it presents the improvement in the form of relative traffic on the SoC fabric (in the vertical axis) against a baseline 8MB L2 cache. Cyberpunk with RT showed a 19% reduction, Black Myth rasterized showed a 36% reduction, and the rasterized Steel Nomad test showed a 17% reduction.

Intel also told us that it has increased its L1 Cache by 33%, noting a move from 192KB to 256KB. When we asked Tom Petersen which area of uplift he thought had the most impact on overall performance, he pointed us toward the register and thread changes. Intel has increased thread count upwards of 25% depending on configuration and has moved to variable register allocation. Petersen noted that occupancy of the compute units (including on Battlemage) previously wasn’t always high, despite them being available for work, meaning that there was more GPU hardware present than was being properly utilized by applications. Intel has focused on this in both drivers and hardware. He noted that previous register allocation and thread count choices would “starve the pipeline if the shader used too many registers,” which is being addressed.

The ray tracing unit also got improvements. Intel says it “slowed down dispatches of new rays while the sorting unit catches up,” citing out-of-order dispatch and triangle testing. The ray tracing unit improvements seem to be largely attributed to asynchronous dispatch-test processes.

Intel also highlighted a new URB manager as part of its fixed function enhancements, which is also where we find the anisotropic filtering uplift. Petersen stated this, “We also now have a new URB manager, which allows partial updates versus flushing the whole thing. Our URB is a structure where we pass results between our units inside of our GPU. It used to be somewhat of a serializing point; now we can actually use that partially without flushing each complex.”

Frame Inspection

We thought these next couple slides were pretty interesting as well:

Intel showed a frame on Xe3 versus Xe2. These are not normalized for configuration size, so it’s not a perfect comparison and it shows a 12-core vs. 8-core configuration, disallowing a perfect like-for-like inspection. This is iso frequency and power, so it is at least normalized there.

The horizontal axis is for API call execution, with the vertical axis being milliseconds of time to execute across a single frame being drawn (higher is worse). This is for Cyberpunk 2077. 

Of note, Intel shows an 8ms reduction to Xe3 with the compute and pixel shader section toward the end, assigning some of that uplift to the change to the variable registers and L1 cache size increase. We can also see that, according to Intel, the L2 benefits the render base pass with a 0.39ms improvement, preceded by the move to 10 threads (and variable registers) providing a 2.93ms improvement in the pre-pass. 

More broadly, Petersen told us in a call that the register allocation and number of threads would starve the pipeline if the shader used too many registers previously, which is being partially addressed here. He said that the previous architecture could cause a reduction in the utilization of available compute resources due to regular flushing of the pipeline due to regular reallocation into memory.

This image is pretty cool and is a look at what actually happens in a frame when it’s being drawn. We have a full video talking about this previously.

Power Delivery

Intel’s focus on power delivery and power management cites learnings from the MSI Claw (read our review) devices and mostly comes in the form of ensuring proper resource allocation for power budget between the CPU and GPU, which should benefit laptop and handheld devices that have a limited power budget split between the two. 

Intel noted that previously, a lack of application awareness meant that the device could sometimes divert too much power to the CPU, leaving the GPU bottlenecked on its power limit while the CPU offered a level of performance that wasn’t being kept-up with by the GPU. 

Intel gave the MSI Claw as an example of a time this didn’t go well.

The company noted that it improved on this earlier in the year with its Intelligent Bias Control v2 and is now introducing a v3 to build upon that. 

Because the system was previously unaware of the application being run, in this case a game, Intel said that software and hardware wouldn’t correctly balance the workload between the CPU and GPU, resulting in stuttering due to being power starved.

“Intelligent Bias Control v2” took GPU heuristics and utilization metrics to then inform thread scheduling and resource assignment at the operating system-level. Intel had previously marketed improvements to 1% and 0.1% low metrics via better frame interval pacing as a result of this change.

The new v3 version of this adds E-core first scheduling, which is self-explanatory in that E-cores get scheduling first when gaming. This sounds worse, and typically would be, but Intel says that the end result is reduced power diversion to the CPU by using lower power cores prior to P-cores, freeing-up more of the shared total power budget to go toward the GPU instead. In GPU-bound scenarios, like many games particularly on handheld devices, this is a better outcome than burning power on a component that isn’t as burdened. 

This comparison between Panther Lake and the prior generation of this bias control solution shows that peaks in power utilization have smoothed-out while the GPU power consumption has leveled to be more predictable. Reminder: This is a first-party tests. The GPU is also getting more total power budget as a percentage than previously, while reducing CPU power in exchange. For GPU-bound scenarios in particular, this should be a better outcome. It might help in some CPU-bound scenarios as well.

XeSS Multi-Frame Generation and Other Changes

Intel also announced XeSS 3, which includes XeSS-Multi-Frame Generation (or XeSS-MFG). A few more letters and they’ll have the whole alphabet. 

XeSS-MFG is conceptually similar to NVIDIA’s MFG. XeSS-MFG takes 2 real frames to calculate optical flow networks using motion vectors and the depth buffer, then uses that information to generate up to 3 frames between the 2 real frames. The frames are then displayed in order and paced in a way to minimize animation error. We also have a separate deep-dive on our new animation error testing methodology.

The new “XeSS Frame Generation Override” setting in the driver software allows the user to set 2x, 3x, or 4x mode.

Intel presented a few timelines of a single frame. One at native, and then several with various levels of XeSS technology in use. The shorter the duration that the frame is on the X-axis, the less time the frame took to complete. The top half of each shows instructions and the bottom half shows when the geometry pipeline is active.

Compared to native, the raster, RT, and denoise sections of the frame are shorter on the XeSS 3 timeline due to rendering at a lower resolution. The first purple section represents XeSS-SR to perform the upscaling. The second purple block starts with the optical flow portion of frame gen, followed by 3 frame generation operations.

It seems like Intel’s argument is that the entire frame gen process takes less time than drawing one real frame, and is therefore better or something, but this totally ignores image quality. We’ve shown with both AMD FMF and NVIDIA MFG that the image quality sacrifice isn’t always worth it. Sometimes it is, but it’s not always as simple as being that way. Intel stated that these frames upscaled with XeSS-SR are the same quality as native, which is unlikely. Intel stated: “That frame is as good as the prior picture, the native frame. But it’s actually being run quicker.” We doubt this will be broadly true and will evaluate later on dGPUs. It was bullshit when NVIDIA claimed it, too. The quality can be good, but is rarely as good.

Intel had some other side-by-sides that we take issue with, and that in combination with still having watermarks on the video means we’ll skip them and just test it ourselves later. 

Intel referred to the frame gen process as looking into the future. NVIDIA CEO Jensen Huang has said similar things about NVIDIA’s frame generation. Both of them are wrong, because all current methods of frame generation rely entirely on finished frames and engine data. These frames already existed and could have been displayed instead of holding them to run the frame generation in between. That isn’t looking into the future, that’s interpolating between two sequential snapshots of the present or near present. Until a predictive method of frame generation comes out, none of these technologies look into or generate “the future,” they at best interpolate the past. And that’s fine, but we’d really like it if these companies could get their shit together and stop saying that they generate the future. 

MFG represented on benchmark charts has been a major and ongoing controversy and misrepresentation of performance on NVIDIA’s side of things. Intel committed to relying on base raster performance without frame generation as the baseline for performance and said that, when it publishes numbers including upscaling or frame gen, those will be provided as supplemental to the base metric. We think this is a better balance of promoting the capability without totally misrepresenting the reality.

Intel also talked about a new version of PresentMon that includes a few changes, partly accounting for frame generation technology.

This article is like our research/whitepaper piece that presents some experiments and possible representations for animation error in benchmarks

The Highlights

• Benchmarking has long had a problem of ensuring numbers relate back to the reality of what players feel when gaming
• Game stutters have sometimes been misattributed to frametime pacing issues rather than the actual problem, which was animation error, aka simulation time error
• The issue mostly comes to explaining precisely why stutters and hitching are happening in games, not just that they exist

Table of Contents

• AutoTOC

Intro

We have a new benchmark metric that exposes a limitation with current GPU and CPU game testing.

Editor's note: This was originally published on October 13, 2025 as a video. This content has been adapted to written format for this article and is unchanged from the original publication.

Credits

Test Lead, Host, Writing

Steve Burke

Testing, Writing, Research

Patrick Lathan

Camera, Video Editing

Vitalii Makhnovets

Video Editing

Tim Phetdara

3D Animation, Editing

Andrew Coleman

Writing, Web Editing

Jimmy Thang

This is that limitation: Frames are displayed at an even pace in this example, but something is still wrong with it. The thing that’s wrong is why we have the new measurement methodology that we’re debuting today: Animation error.

Animation error is the difference between the pacing of animation and display. Putting "animation" in the name might be confusing: Intel called it Simulation Time Error at one point. 

And this is a sample chart illustrating what we’re talking about, where you can see the animation error timing sometimes in alignment with frametime spikes, sometimes out of alignment with it, showing that it’s a different thing. 

Here’s another chart showing the percent animation error during a test pass. 

Or this one, where we show the CPUStartTime versus animation time delta.

These are all new types of benchmark charts that haven’t been shown before. Tom Petersen first pitched the idea of animation error a couple years ago, something he arguably began on about 14 years ago.

The problem with GPU, CPU, and new game benchmarks has always been that it’s tough to accurately capture the actual player experience. Framerate was a good start, and frame-to-frame interval testing (or "frametime testing") was a great expansion on that -- but neither perfectly captures the real experience.

This is what we’re unveiling and detailing today. Consider this a whitepaper, like a research piece that’s intended to put information out to the community for people to start trying to experiment with. None of this is perfect yet, but we think we have a good foundation for viewers and other reviewers to build upon and advance our understanding of game behavior. If you’re a reviewer and this is useful, please point back to this story as it was over a month of work for us to wrap our heads around.

Let’s get into it.

This methodological deep-dive lays the foundation for new testing. It’s exploratory.

We rolled-out 1% low and 0.1% lows in our testing back in 2014, eventually popularizing their presentation on bar charts alongside average framerate. These are the metrics we use today to point us toward a problem with frametime pacing. The industry has relied on frametimes, 1%, and 0.1% averages for over a decade now with few new metrics in between. 

The history to that is all important, and it’s important that those who did the groundwork before us are known: Tom Petersen, PC Per’s Ryan Shrout, and Tech Report’s Scott Wasson all advanced this metric, with Petersen doing heavy lifting on providing software tools and early insights to frametime analysis. His engineering work has continued with the open source tool PresentMon, which gets us to animation error today.

Animation Error should be thought of in a more traditional sense: Like a flipbook with perfectly animated drawings, but without the execution of flipping through the pages at a constant tempo. That's just one specific scenario for animation error; you could extend the metaphor, like flipping through the pages perfectly but messing up the drawing (or you could do a combination of both).

We're talking about animation in the dictionary sense, "a movie, scene, or sequence that simulates movement from a series of still frames," so animation error applies to entire frames. Think of it like frames of drawings on an animation lamp, in a flipbook, or on a reel of film. We're NOT talking about errors in animation of individual models, objects, NPCs, or sprites within the frames, and we’re also not talking about games that just have bad animation from the artists. 

You could technically have animation error even staring at a blank wall in-game without any movement whatsoever, though it might be impossible to notice.

This all comes down to two things: Smoothness and acceleration. When we were talking to Tom Petersen about this concept, he made some good points about this. Showing frames faster allows the brain to interpolate and generate an illusion of smooth motion, which is the illusion of TV and movies. But the brain also knows how to identify acceleration, something Tom equated back to “monkey times like running away from lions and shit.”

This is an example of animation error that we created in 3D space. It’s very smooth and accurate as a rate, but it’s slow. Every now and then, you’ll see an error in it despite the smoothness of the frames, and that’s stutter. When the brain sees even subtle acceleration or deceleration, we pick it up quick. That’s what makes it feel so bad when we see stutter in gaming. Stutter is what you’re seeing here: Something accelerates or decelerates quickly and the body is overreacting because, to quote Tom, “we don’t want to be eaten by a lion, or some shit like that.” He really has a gift with words.

But it would be great if we could measure smoothness and acceleration separately, because they’re different problems: Framerate and frame-to-frame interval evaluate smoothness, but acceleration is something we haven’t done a good job at quantifying in this industry. That’s characterized by animation error, which we’re introducing with its first full charts today.

This chart is from our Dragon’s Dogma 2 testing at launch, where we were quietly beginning to farm data for this eventual piece. You can see where the animation error blips often align with the frametime spikes, but not always.

AnimationError can also be positive or negative. When a frame is displayed "too late" (relative to its correct placement) that's a negative error, and when it's shown "too soon" that's a positive error. Neither is good. At a macro level, it doesn't matter which is which: further away from 0 is always worse. 

The idea is that if frames are created at a certain pace, you should see those frames displayed at the same pace. There are differences between animation error and frametime pacing, though. 

If there's a mismatch, that's where the animation part comes in: movement depicted in the frames will appear jerky and wrong, even if frametimes are perfectly consistent. 

For a real-world demonstration, the simplest, most reliable way we found to directly induce animation error was with SLI. 

Yeah, we know.

We can take one card and get a normal result, then add a second and get a result with higher animation error. 

That limits us to older hardware, and it also limits us to GPUs that we own in pairs. We selected two 1080 Tis (read our revisit) since they're our newest cards that still use regular old SLI, and (as of now) they're still supported in the most recent NVIDIA driver package. We also had to select a game that supported SLI.

We’re getting into benchmarks. We’ll break these down starting at the most abstracted metric, which is this chart. 

This is framerate represented as bars, abstracting away from time. Next, we’ll look at the frametimes that create this framerate average, and last, we’ll look at the new animation error metric.

These are averaged results for 30 second logs of Far Cry 5's baked-in benchmark, one using a single 1080 Ti and one using two 1080 Tis in SLI. The 1% and 0.1% lows indicate that there weren't huge frametime spikes, and it’s these metrics that tell us when we should inspect a frametime plot closer for major problems. Average FPS smooths over problems, 1% and 0.1% are still averages and can still smooth over them, but are more likely to draw our attention toward a problem because they’re averaging the worst 1% and worst .1% of data.

As a reminder, we aren't using percentiles, which is a different way of approaching this. We explained that in our video where we did JayztwoCents' lab overhaul.

The average with SLI is higher, as expected; however, beyond that, there’s no dramatic change between how these numbers manifest. The single GTX 1080 Ti appears to have closer frametime pacing to its average, which is what we’ve been preaching for years as a good result, but the dual 1080 Tis still look good overall.

Data Presentation: Frametimes

The frametime plot helps us see deeper into those bars. 

This still isn’t animation error, though, and it’s still not new.

This plot of frametimes is for two individual passes. As we know based on the last chart, the SLI run’s gap between the average and its 1% and 0.1% metrics is wider. Here, that materializes in the form of spiky behavior (particularly in the 500-1000 frame range) for the SLI configuration. The experience is far less consistent, with more sporadic frametime excursions from baseline. Most users begin to notice these around 8ms, according to an interview we conducted with Scott Wasson years ago, but only if they frequently occur. 

The single-card run has more consistent frametimes, despite its lower average. It’s not so bad that it’s a ruinous experience, as the SLI configuration is ultimately still within the range of 2-4 ms of the baseline, but the relative distance from baseline is larger. It’s not like we’re seeing 100ms spikes where you’d stare at one frame for 1/10 of a second, as we’ve seen in other tests on modern single cards.

Data Presentation: Animation Error

Here’s the new stuff.

The left axis shows animation error in ms, with deviations away from zero depending on whether frames showed up too soon or too late. 0 is perfect and 0 does occur. The X-axis represents frames, the same as the frametime plot we just showed. The animation errors could also be taken as absolute values for a lower-is-better representation, but this is a plot of the raw data as logged by PresentMon. 

There are many ways this data could be plotted. A scatterplot lets us judge the data points individually, but still keeps the points in order so we can see how behavior changes over the course of the test. Lines between the points wouldn't mean much here.

By plotting the animation error, we can get rid of frametimes as a variable and just compare that relative spikeyness as a player would truly feel it interacting with the game itself. This is closer to the real experience, in the same way end-to-end latency can be but for different reasons. The further the deviation from zero in either direction, the worse. The cards in SLI frequently had 2-3ms of animation error per frame, while the single card typically had well under 1ms animation error. The SLI configuration is significantly worse for animation error in a relative sense, despite even the frametime data looking not that dramatic.

The single card is clearly far better in terms of animation error, although we need more data to judge whether the SLI result is "bad."

Strange Brigade Animation Error

This is another test. This time, we’re using Strange Brigade. Our goal today is exploratory, so we’re choosing games based on usefulness to explore the concept, not on their popularity. That’ll come later.

The main advantage of a plot like this is that we don't have to insert our own calculations or conclusions: we can simply show the data. The single 1080 Ti sticks even closer to zero animation error per frame than it did in Far Cry, while the SLI 1080 Tis continue to generate 2-3ms of positive or negative error on nearly every frame, only rarely approaching zero. 

SLI outperforms the single-card so heavily in this title that the red line is significantly shorter, which is a downside of using frames as an X-axis. Even though both its average framerate and its 1% and 0.1% lows are overall good here, the animation error is far superior on the single GPU. There is a possibility that it feels better to a player, but not for the reasons everyone in this community has said for years: It’s not due to frametimes, on a technicality, but animation error, which is a metric that has been under the surface this whole time.

Animation Error Bar Chart

This is an “Error Per Frame” chart we attempted, which puts the data back into bars for denser comparison of more cards. Maybe this could work better as a visualization: for those two individual passes in Far Cry from earlier, the total animation error (taken as absolute values) divided by the total number of frames was 0.13ms per frame for the single 1080 Ti and 2.31ms per frame for the SLI 1080 Tis. These are basically single-number summaries for the plots we just showed. But now we’re doing the bad thing again: We’re abstracting away from the base metric over time (shown as frames) and converting it into a bar, because that’s easier to read. This isn’t ideal for a lot of reasons.

First of all, this could obscure individual big error spikes (if that's something we want to track), and secondly, it also adds framerate back in as a variable. The SLI setup generates more frames, which lowers the end result for this calculation, arguably making the SLI cards look unfairly good; then again, maybe it’s not “unfairly” because more frames going by quicker could help disguise animation error.

This is getting complicated, but you can see why we’ve had to think about this for weeks.

Far Cry 5 Animation Error Percent

So we don’t like the prior chart for those reasons, and the scatter plot doesn’t accommodate more than two or three GPUs before it’s illegible. Maybe this will help.

This alternative was suggested by Tom last year. It divides the total animation error absolute values by the total frametimes (the length of the test run) to get a ratio or percentage. 

This is equivalent to comparing those average error-per-frame numbers from the last chart to average frametimes. 

However, in order to make this chart we’re showing accurate and not misleading, we had to use two times the total frametimes for the calculation because otherwise, the implication is that 100% is the maximum value, and it wouldn’t be if we hadn’t corrected for that, and that’s because in the absolute worst case scenario and assuming that latency doesn't accumulate over the course of the test run (which is a different subject entirely), the maximum total animation error would be twice total frametime. 

THIS diagram makes that abundantly clear by showing that alternating between infinitely small frametimes rounded to zero and 10ms frametimes, and inversing the display times, we can get 10ms or -10ms of animation error per frame, where adding up all frametimes gives us 30ms yet adding up absolute values of the animation error gives us 60ms.

OK, this graphic really isn’t helping make it less confusing. The point is, we already corrected for this in our percent chart and any other reviewers planning to use such a chart will need to do the same.

We used the same two Far Cry passes for this example: for the single 1080 Ti, the result is 0.7%, and for the SLI setup the result is 17.3%. This potentially cancels out or reverses the high-framerate advantage from the last chart: more frames equals more (total) error.

Real-World Uses

Let’s talk potential real-world use cases. Maybe this will give other reviewers some ideas.

Animation error is theoretically decoupled from framerate and frametime consistency, although in reality poor performance correlates across all of those categories. 

In this mockup from Intel, the bar marked "CPU Stutter" marks a frametime spike, while the mismatch between the sizes of the bars in the top and bottom rows is the animation error (F1 versus F1 = error for F2, F2 versus F2 = error for F3), so we'd see similar spikes in both metrics.

Animation error is also separate from latency: it's about how the frames are spaced, not whether they're all showing up ten seconds late.

Frame generation has interesting implications for animation error, but unfortunately there's no animation time for fake frames, so there's no reference point for calculating error. That makes this whole thing even more challenging to quantify.

We'd like to check on NVIDIA MFG in particular, since that comes along with flip metering that shuffles frame timings around at the end of the pipeline, which has the potential to actually induce animation error (as we mentioned with the SLI example).

Animation error could shine a light on shortcomings of flip metered frame generation that are currently masked with existing testing methods, but we won’t know until the software works on it, which may at least partially rely on NVIDIA’s willingness to play ball.

Dragon’s Dogma 2 Animation Error

Because we expect AnimationError to correlate with frametime spikes most of the time, its most common usages for us will be similar to 0.1% lows and frametime plots, but with a more direct representation of how the game feels. 

For example, this is a chart that Tom helped us generate during the early stages of troubleshooting Dragon's Dogma 2's performance back when it launched -- that’s how long we’ve been thinking about this data. 

The spikes in both metrics line up exactly, but animation error adds depth by telling us why the stuttering was so noticeable and unpleasant. The positive and negative animation error dots that correspond to the frametime spike toward 60ms around frame 500 shows a potential for 45ms of animation error. That’s potentially a lot more noticeable than even the already noticeable frametime hitch of 58ms.

Borderlands 2 Animation Error

Here’s Borderlands 2. Yes, we know there are newer ones. But this is a better demonstration.

Since animation error effectively cancels out framerate differences, we can use it to compare two completely different pieces of hardware and get more nuance out of it.

These two results are from our piece about the death of 32-bit PhysX. These two particular results were fairly close together both in terms of average FPS and lows: the GTX 980 with GPU PhysX averaged 101 FPS across multiple passes and the RTX 5080 (read our review) with CPU PhysX averaged 95 FPS. These are real results and we explained them in our piece about NVIDIA killing 32-bit PhysX support this year.

In spite of that, there's a clear difference in behavior, and animation error is much higher with the GTX 980. Using the percentage math we mentioned earlier, that's an error-to-frametime ratio of 11.9% for the 980 and 1.8% for the 5080, which allows us to identify a problem without having to generate frametime plots for every single result.

Percentile Limitations

We probably wouldn't use a test scene that resulted in a graph like the GDC 2015 one shown above by NVIDIA, but as it points out, there are scenarios like this where frametime spikes could be artificially masked and not show up in 1% and 0.1% low calculations. 

NVIDIA is using percentiles here instead, which we don’t use, but the idea is similar. 

Assuming these frametime spikes were accompanied by animation error, calculating animation error would do a much better job of summarizing the problem (in this instance). That said, our approach to lows already helps to control for some of this, but it still requires knowledgeable testers to know when to look into the 0.1% and 1% low results. 

Microstutter & Multi-GPU

Animation error has been associated with microstutter in the past, but it's not quite the same thing. If anything, animation error is a way to measure microstutter, but not microstutter itself, depending on the definition. Microstuttering was frequently brought up in the context of multi-GPU rendering, so we'll go back in time and start there by referencing materials from GDC 2015.

This simplified timeline establishes our foundation: each block is a frame, and frame N gets displayed while the computer works on frame N + 1 behind the scenes. CPU work isn't represented here, but we'll ignore that for now. The frame times are perfectly consistent, which is ideal. 

This next slide shows how Alternate Frame Rendering (AFR) multi-GPU operates, or at least operated back when anyone actually supported it. Each of the GPUs takes turns rendering frames, and the output is combined and displayed in order. The blue boxes are still uniform sizes, indicating consistent display times, and they're smaller than the previous diagram, indicating a higher framerate.

One of the major difficulties with AFR is trying to synchronize GPUs. Here, GPU0 and GPU1 are taking the same amount of time to complete individual frames, but they're poorly synced so that some frames get little display time and are effectively wasted. 

These are also known as “runt frames,” where fractions of frames are shown in a way that elevates the average FPS, but creates an awful experience with bad tearing. The average framerate is higher because more frames were technically shown, but the additional frame is useless and the actual experience can look lower in framerate, which is one definition of microstutter.

Again, we can't see the CPU stage of the pipeline here, but we'll assume it's being completely consistent. 

In this example, the rhythm of animation matches the rhythm of display, so there is no animation error. The framerate is stuttery, but all the moving stuff in the frames shows up in the right place at the right time. After each short frametime, objects move a little; after each long frametime, objects move a lot.

You can avoid that microstutter by forcing the pacing into alignment. Ideally this happens by manipulating delays early in the pipeline (on the CPU), in which case you return to a clean result like this one.

However, if you were to meter those frames out at the END of the pipeline, that would directly contribute to animation error (or "animation stutter," as NVIDIA put it back then). We've created and rendered an edited diagram to show what that would look like. If you just take the "short" frames and hold onto them longer before flipping, then there's a mismatch between the pacing of the frames as they're displayed versus the animations depicted in those frames. That leads to perceived stuttering and rubberbanding.

Back to our earlier animation, that’s seen when comparing the red and green indicators below the scene, where the imperfect ball stalls and then gets dragged forward in uneven intervals.

We aren't exploring which method NVIDIA or AMD used to deal with microstutter; that's a subject for another time, and that time was 12 years ago. Today, we're just showing a real world example of a situation where animation error was a risk.

Animated Examples

Andrew on the team made some 3D mockups of simulated animation error. We started by rendering out a scene at 120 frames per second. This could have been any arbitrary number, but picking a high framerate allowed us to downsample and play around with the spare frames.

Here's what that looks like. The top row of squares represents our 120 FPS source video, and we'll pretend that our simulated game has an impossibly low latency of zero, meaning that this row represents in-game reality. The bottom row of squares represents frames that we pulled out of the original sequence to create our 60 FPS video. Because we pulled exactly every other frame, we still have smooth playback with zero frametime spikes and zero animation error. Because we display the frames in sync with their original placement, we have zero latency as well.

Since we're pulling frames every 16.67ms from the source video, our AnimationTimes are always 16.67ms. And since we're also displaying those frames every 16.67ms, our DisplayedTimes are always 16.67ms. Therefore, AnimationError is always zero.

The green circle represents "reality" as determined by our source video, while the red circle represents what we're actually seeing. Again, these match perfectly in this control example.

By taking some of the spare frames from the source video, we can create animation error, but we need to be specific. The diagonal lines mean that we're taking the original frames and displaying them later than "reality," which introduces latency. Latency itself is not animation error. That’s a different problem. 

As we play through the big lump of diagonal lines at the start of this clip, the red circle falls behind the green circle, but the animation of the video remains smooth. Animation error is when the red circle jerks around, skipping to catch up with the green circle at an uneven pace. If you watch the video clip when this happens, you can see the interruptions. This is what we mean when we say animation error is a measurement of jitteryness.

Here's an extreme example of our point about latency: we're displaying every frame 41.67ms "late," so the red circle lags behind the green, but the resulting video is identical to the control. AnimationTimes and DisplayedTimes are still perfectly matched 16.67ms intervals every time, so there's zero animation error.

We can do this multiple times during the clip to intensify that feeling of rubber-banding. We're creating a lot of variation in our simulated AnimationTimes here: if we take two back-to-back frames from the 120 FPS source video, that's an 8.33ms AnimationTime. If we pull two frames that were spaced four apart in the original video, that's a 33.3ms AnimationTime. Meanwhile, our DisplayedTimes remain constant, because we're still displaying fresh new frames exactly 16.67ms apart. That's the mismatch in pacing that animation error quantifies.

As Tom told us before, our challenge is to "make people understand that you can take frames and show them with an even cadence on display and still have it look like shit." He has a real way with words. It’s like poetry.

We can also create an inverse example. Think about a flip book. Here we've pulled frames from the source video at even intervals for constant simulated AnimationTimes of 33.33ms, but by displaying those frames at uneven intervals, we still create animation error. This is like drawing a perfect flipbook animation and then failing to flip through the pages at a constant tempo. This is a weird theoretical example, because if we assume this is a case where frametimes equal animation times, PresentMon would report a completely steady 30FPS based on MsBetweenAppStart.

That's not the only way for animation error to manifest, though—in fact, it's pretty unlikely that you'd naturally encounter perfectly consistent DisplayedTimes with inconsistent AnimationTimes, or vice versa. A more realistic scenario is this one, which simulates CPU-based stuttering, like the diagram Intel shared with us.

Here, rather than displaying a unique frame every 16.67ms, we freeze on individual frames. These are DisplayedTime spikes, which usually correlate to FrameTime spikes under the hood.

For each spike, we get two animation errors: after a frozen frame, the next frame is judged to be too late, and the frame after that is judged to be too soon.

This is closer to what we've observed in games, but just like with latency, it's important to remember that the frametime spikes are not the same thing as animation error: you could theoretically keep freezing on frames while maintaining zero animation error. 

To help explain, let's cover a real-world example.

Capture Demonstrations

For easier discussion, we're mostly ignoring VSYNC and variable refresh so that the monitor is not a factor in any way. 

When we say frames are "displayed," we mean that a flip has been signalled to the operating system, and without VSYNC, that flip can happen even if the monitor is in the middle of a refresh (leading to tearing). 

That pushes numbers logged without VSYNC towards the theoretical realm, but no more so than usual: for example, in our launch review, the RTX 5090 averaged a ridiculous 407 FPS in the Dawntrail benchmark at 1080p. That's a comparable performance number, independent of whatever monitor we used, and in that context, that's what we wanted because we want percent scaling between devices.

Separately, higher framerates do correlate with lower latency, so there’s value from that side as well. Today though, we’re also ignoring latency for purposes of focusing discussion.

All of that said, the easiest way to actually show animation error in captured footage is with VSYNC to avoid tearing. If you ignore the overall drops in framerate and focus on the movement of objects that should be smoothly traveling across the screen, you'll see them appear to change speed and jump around: that's animation error.

It's most noticeable in fast panning shots with smooth tracking: the camera should be moving at a steady rate even when the framerate drops, but it appears to hitch and rubber-band, especially when played back at half speed. Animation error is separate from frametime spikes, but the two things are frequently associated, and they're both bad.

We want to be careful here, because animation error is uniquely bad when VSYNC is enabled, and this footage isn't representative of the non-VSYNCed test passes that we're about to discuss.

One of the only ways we can represent those test passes in fixed-framerate capture is with an FCAT-style overlay, which adds a visual indicator of where torn frames begin and end. This helps illustrate runt frames as well, and was used in VirtualDub back in the day. That gives us an indicator of each individual frame that we're discussing without adding the complication of VSYNC, but it does also mean that frames may only show up as a tiny sliver of pixels.

If we play back this footage slowly, you can frequently see the pattern of tearing: a new color shows up at the bottom of the bar in one frame of the capture, then it continues from the top of the bar in the next frame of capture. It's kind of a mess, and it's difficult to focus on one sliver of frame at a time, hence using VSYNC for visualization in spite of the downsides. The FCAT functionality would be more useful as part of the PresentMon overlay, which can simultaneously show a live graph of animation error.

Advanced Definitions

Here are the PresentMon CPU metrics.

PresentMon CPU Metrics:

CPUStartTimeInMs: The moment where a new frame is born, expressed as a timestamp (relative to the beginning of the PresentMon session).

AnimationTime: This is PresentMon's best estimate of the frozen moment in game time that's depicted by a frame. As a player, you're always seeing rendered images of a game several milliseconds after the in-game reality that they depict. AnimationTime is the timestamp for that reality. 

According to Tom Petersen last year, "Today people are mostly using CPUStart as the AnimationTime, which is a pretty good proxy, and that's what we're going to be doing initially. There are explicit APIs, both from NVIDIA and from us [Intel] and others that are allowing game engines or games to tell you that AnimationTime. And so as that becomes more available, we'll be building that into PresentMon." 

Logically, AnimationTime should be the same as the moment the frame was born (CPUStartTimeInMs), but games can pull tricks to smooth animations so that AnimationTime for a frame doesn't line up with wall-clock time. Under normal circumstances it should be close enough, though.

As an example of an exception, PresentMon can monitor SimStart events when using Intel XeLL (and soon NVIDIA Reflex) and base AnimationTime on that instead. That's valuable because Reflex and XeLL clear the render queue and keep the CPU sitting around waiting for input until the last possible second, so there's a higher potential for differences between CPUStart and the true animation time.

Here we have a PresentMon capture of Cyberpunk 2077. The X-axis shows individual, logged frames, and the Y-axis is the delta between CPUStartTimeInMs and AnimationTime for each frame. Normally, this would result in a perfectly flat line at zero, but since we have XeLL enabled, AnimationTime is based on SimStart instead. With XeLL there's a significant delta between the two values on nearly every frame, which shows that AnimationError would be incorrect if it were based on CPUStartTimeInMs when low-latency modes like XeLL, Reflex, and Anti-Lag 2 are enabled. AMD's Anti-Lag 2 doesn't generate SimStart events that PresentMon can grab, so (for now) we won't be able to accurately score AnimationError with that feature enabled.

MsCPUBusy: This period begins at CPUStartTimeInMS and includes steps that Intel labels as Game and Render. "Game" is the time spent handling game logic and calculations for the frame, and "Render" is the time spent converting the results into API calls (DirectX, Vulkan, etc.).

The end of these CPU-specific tasks is marked by the Present() call, which signals to the GPU that it has everything it needs for rendering. Future versions of PresentMon may break this down further because CPU work is complex.

TimeInMs: This is the timestamp of the Present() call we just mentioned. It's important to remember that this call doesn't mark the beginning of the GPU rendering step, because the GPU can get a head start before the CPU is done generating API calls. 

Usually the end of the Present() call is the CPUStartTimeInMS of the next frame.

MsBetweenPresents: The delta between this frame's Present() call (TimeInMs) and the previous frame's. In the old days, time between Present() calls was used as a (fairly good) approximation for frametimes, but it's technically a different thing. For that reason, MsBetweenPresents is unusable for per-frame calculations like animation error.

MsInPresentAPI: This is the same as MsCPUWait. This is the period between the Present() call and the moment when the CPU begins working on a new frame, meaning that there's nothing blocking further CPU work. 

MsBetweenAppStart: PresentMon's best representation of the literal time taken to create an individual frame start-to-finish (an improvement over MsBetweenPresents). It's the delta between the CPU starting work on one frame and the next, so the difference between CPUStartTimeInMs for the current frame and CPUStartTimeInMs for the next frame (or MsCPUBusy plus MsCPUWait).

MsBetweenSimulationStart: This column would depend on SimStart events from Reflex or XeLL. In the current version of PresentMon, MsBetweenSimulationStart is "disabled until underlying event support is enabled."

PresentMon GPU Metrics: We don't need these numbers in order to calculate AnimationError, but we'll go over them briefly. They include:

MsGPUTime: The total GPU render period, comprising GPUBusy and GPUWait periods. This was formerly called msGPUActive.

MsGPUBusy: The portion of the render period "during which at least one GPU engine is executing work from the target process."

MsGPUWait: The portion of the GPU render period where the GPU was idle, potentially due to some codependency on CPU resources.

PresentMon Display Metrics: 

MsUntilDisplayed: The time between the Present() call for the frame (TimeInMs) and the time at which the frame is displayed. You can calculate the timestamp at which the frame is displayed by adding these numbers, but it isn't logged directly. "Displayed" here means that a flip (pointing to a new frame buffer) is signalled to the operating system. This is different from new pixels literally showing up on the physical monitor, although the timing should be very close.

MsBetweenDisplayChange: How long the previous frame was displayed before the current frame started to be displayed. There's an argument to be made that this reflects the user experience more directly than MsBetweenAppStart, but MsBetweenAppStart is directly tied to performance, so that metric is still better for testing hardware. However, since display happens at the end of the pipeline, MsBetweenDisplayChange is the only way to include post-processing stuff like generated frames and RTX 5000 frame metering in results (if you want that). 

Combined Metrics:

MsGPULatency: The period between the absolute start of work on the frame (CPUStartTimeInMs) and the point at which the GPU started working on it. The start of the GPU render period can be inferred from this.

MsRenderPresentLatency: This is the period from the Present() call at the end of CPU rendering to the end of GPU rendering. This is equal to MsUntilDisplayed unless VSYNC is enabled.

MsAnimationError: Here's how Intel represents the animation error formula for frame N:

(AnimationTime_N – AnimationTime_N-1) – MsBetweenDisplayChange_N

Again, the result of the formula can be positive or negative, but further away from zero is always worse. Adding together all the positive and negative animation errors for a logging period will typically cancel out, so to get a useful total we need to take absolute values.

In the words of Tom Petersen, "The animation step is basically equal to the frametime, mostly. There's some times where it's a little different. But what you need that for is to be correlated with the DisplayedTime step. Because if the DisplayedTime step is different from the animation time step, you'll get a simulation time error, which is measuring stutter directly for the first time."

We've established that AnimationTime is the in-game point in time that a given frame depicts. The delta between animation times for consecutive frames is the amount of in-game time that has passed between them.

We've also established that MsBetweenDisplayChange is the time that a frame is displayed before the next one shows up.

If a long time passes between taking snapshots of the game state, a long time should pass between displaying the snapshots. Even if the AnimationTimes are spiky and uneven, the DisplayedTimes should be matched exactly, or else you get AnimationError.

Conclusion

This doesn’t replace current testing or run instead of it. It’s another tool -- similar to frametime charts -- to help better understand what’s happening in a game. There’s also a lot of theoretical situations here, so it isn’t always practical.

1% and 0.1% lows as bars on a chart took on a life of their own over the last decade. They are still the fastest “glanceability,” and we’re glad we introduced them to our charts now 11 years ago, and we’re going to continue to use them. But it’s time to try and find new metrics, and we hope animation error can supplement the 1% and 0.1% average bar representations of frametime pacing as another means to determine why a game just feels bad sometimes.

But we don't want to oversell what this number actually means. In Tom Petersen’s words: this is a "how jittery am I" metric. Our work here on and off over the past couple years, and more seriously over the past month, has been trying to prove the concept and find a way to put it on a chart that makes sense to anybody. 

We devoted a lot of time to explaining how and why it's different from the numbers that we already measure, but in practice and in most cases, we expect it to complement those numbers, not contradict them. 

That also means we aren't necessarily expecting any huge upsets versus what we've already concluded in existing reviews.

That said, this is a valuable new tool that can do several other things for us: it can show us when we need to make a frametime plot more easily, it can show us when stuttering happens independently from frametime spikes (although that may be unlikely), it can (sort of) normalize for frametimes in a way that makes comparisons between different hardware easier, and it can deal with up-and-down frametime trends across test passes effectively.

Most importantly, animation error forces us to think about why we measure the things we do. We're now closer to discussing why stuttering feels bad, not just the fact that it exists. 

We're still experimenting with ways to make it useful. 

If you want to try it out for yourself, PresentMon is free and open-source, and it now has a GUI version available as well. If you’re a reviewer and you find this useful in developing your own methods, we’d appreciate you pointing back here.

We’ve been using PresentMon for years, and actually, most people who’ve tested game performance have -- they just often don’t know it. 

PresentMon is wrapped by half a dozen other tools that reskin it or use it in some capacity and it’s an open source project with contributions from around the hardware community.

We use the command line version, but there’s also a user interface tool that you can see in our video where we introduced Jay to it previously. They sometimes have different features.

Because of that, a quick security warning first: do not visit PresentMon dot com or download anything from that site. PresentMon is hosted on GitHub and Intel.com and is an open source utility. Usually when someone pretends PresentMon is their own project and reskins it with an interface, they at least come up with a new name. For security purposes, we’d advise only downloading PresentMon itself from GitHub or the Intel site.

Testing animation error like this is exciting because we can finally directly score stuttering instead of simply deducing it. Big picture, this is similar to when PresentMon moved away from MsBetweenPresents: the numbers and conclusions may not change much, but the measurements are closer to what we're really talking about. 

This isn’t the endgame of benchmarking. Hopefully there won’t be one, because that’d be boring. There’s a lot more to learn and this is exploratory and just us putting research out to the internet to experiment with. It’s also up to the vendors to play ball with open source tools like PresentMon. Experiment with the new ideas and point back to us if you find our work helpful as a foundation, and credit to Tom Petersen for opening up the tools to measure these metrics.

We benchmarked the RTX PRO 6000 in gaming scenarios using an AMD Ryzen 9800X3D, experimented with LLM benchmarks, and ran thermal and acoustic tests

The Highlights

• The RTX 6000 features 24,064 CUDA cores, which is nearly an 11% increase over the RTX 5090
• Unlike the 5090, which uses liquid metal, the RTX Pro 6000 uses thermal paste
• On the gaming side, the RTX Pro 6000 outperformed the 5090 by roughly 5 to 14% in our tests, but that alone isn’t a good reason to buy the card
• Original MSRP: $8,000 - $11,000
• Release Date: March 18, 2025

Table of Contents

• AutoTOC

Intro

We paid $8,000 for an RTX PRO 6000 GPU with 96GB of VRAM when we bought it from Brent Rambo, the actual guy in this meme, who’s grown up now and had access to an RTX Pro 6000. These days, Rambo is busy with custom and bespoke PC builds these days on his own website, and that means he has access to interesting hardware.

It’s a shame that NVIDIA has left us feeling like it’s holding its engineers hostage for manipulating reviews, but it’s clear that we don’t mind buying our own hardware to review.

Editor's note: This was originally published on June 24, 2025 as a video. This content has been adapted to written format for this article and is unchanged from the original publication.

Credits

Test Lead, Host, Writing, Editing

Steve Burke

Testing

Patrick Lathan

Video Editing

Vitalii Makhnovets

Camera, Video Editing

Tim Phetdara

Testing, Writing, Camera

Tannen Williams

Writing, Web Editing

Jimmy Thang

The RTX PRO 6000 is the best marketing for AMD’s Ryzen 9800X3D (read our review) that we use in our GPU test benches, mostly because we’re seeing 5% to 14% scaling in gaming benchmarks versus the RTX 5090 (read our review). We also tore it down and found the most insanely dense PCB we’ve ever seen with a reversion away from liquid metal.

But this doesn’t have 96GB of VRAM just because that’s where it all went when they lost it on the RTX 5080 (read our review), 5070 Ti (read our review), 5070 (read our review), 5060 Ti 8GB (read our review), and 5060 (read our review)-- it’s also there because it’s better for... AI.

Today, we’re benchmarking some of those in a round of experimental benchmarks. We haven’t done machine learning and LLM tests before and we don’t have a full methodology defined yet, but we’ve begun experimenting.

NVIDIA RTX 6000 Overview

This card is not for gaming. That much is obvious. We’re still testing it in games, and that’s mostly for our own purposes: We want to know how much room the 9800X3D has in it to scale for GPUs, and fortunately, it looks like there’s still plenty for another generation of reviews. The 9800X3D keeps up with the RTX PRO 6000 even when at 1080p in some situations. 

The biggest difference between the PRO 6000 and the 5090 after the 96GB vs. 32GB memory capacity is the GPU itself: The PRO 6000 has 24,064 CUDA cores to the 21,760 of the 5090, nearly an 11% increase. That’s a big difference.

There are multiple versions of the RTX 6000. The one we're looking at specifically is called the Blackwell Workstation Edition, but NVIDIA also has so-called “server editions” of the cards, and those are intended to go into servers and racks where they're getting force fed air from the front of the chassis rather than using an FE style cooler, which is the one we're looking at today. So, the one we're looking at is more similar to what we would see in sort of the consumer to prosumer workstation class.

NVIDIA RTX 6000 Specs, Architecture Basics, and Price

Getting into the specs:

The RTX PRO 6000 Blackwell GPU is built on the GB202 Blackwell die. The full GB202 die has support for up to 24,576 CUDA cores, 768 TMUs, 192 ROPs that might even be present, and is also used for the RTX 5090.

The PRO 6000 variation on this GPU has 24,064 CUDA cores, so it’s down by 512 and isn’t a perfect die, with 752 TMUs (down from 768) and still 192 ROPs, and they’d better all be there for the price. The card is also running 188 RT cores and supports PCIe 5.0 x16. Advertised clocks are 1590 MHz base and 2617 MHz boost. The real reason people buy this card is its memory, though, at 96GB of GDDR7.

The closest consumer class card to compare against is the RTX 5090. It’s built on the same die, but with 176 ROPs, 680 TMUs, 170 RT cores, 21,760 CUDA cores, 32GB of GDDR7 memory, a 512-bit bus width, and a far lower price at $2,000 to, realistically, $3,000 rather than $8,000 to $11,000.

For some notable differences: The 5090 only has a third of the memory capacity of the PRO card, and the RTX PRO 6000’s die is almost 98% present while the RTX 5090’s die is closer to 88.5% present, with the remainder either being fused off or defective fallout and fused off. 

You could also technically buy four 5090s for the price of one RTX PRO 6000, so the use case really does come down to having as much memory as possible attached to a single GPU and PCIe slot.

NVIDIA RTX 6000 Tear-Down

The RTX Pro 6000 we tested is a 2-slot card with 2 large fans on the front. The PCB is in the center of the card and you can see where the fins start to get taller. This indicates where the PCB plate ends. 

The card’s PCIe slot is connected via a customized pinout adapter. 

If you look at the backside of the card where the fans are, you’ll notice a slight depression down towards the middle of the fins. This is supposed to help with pressure drop. The design is full flow-through, which pushes air straight through the card rather than into a PCB. 

The other side of the card looks very similar to a 5090. 

The top of the RTX Pro 6000 retains the angle on the card’s vented slats. This projects the air out at an angle that’s up and away from the card. 

Like the 5090, the card uses a 12VHPWR cable.

Starting disassembly, we removed the card’s back plate. Then we removed a single screw, which allowed us to pull off the card’s exterior frame.  

From here, we removed some screws around the center of the card to remove the center covers. Doing this reminded us of how masterful the mechanics of the FE design is. It’s assembled really well, it’s easy to take apart, and there’s no mechanisms to try and prevent disassembly.   

Next, we unscrewed the clamp that holds the PCIe slot to the card.

Here’s a close look at the removed PCIe adapter. 

From here, we pried a plate off of the card, exposing the PCB. 

This gave us our first look at where the memory is connected. We counted 16 memory modules on the back with the clay type of thermal pads. 

The PCB is crazy dense, which is an engineering challenge. 

We then proceeded to remove ribbon cables and screws to the leaf spring to access the GPU. 

Pulling off the PCB, we were expecting to see liquid metal, but we saw thermal paste instead. We really weren’t expecting that. We also saw a massive GPU substrate.

Upon removing the PCB, we noticed that 1 pad went over the edge a bit. 

Analyzing the paste application pattern, it looked like the heaviest imprint was right in the center with lighter pressure on the outer edges, which is fine.  

Looking at the thermal pads, they were making clear contact, though we did see some pads that were smashed over the sides a bit. This is fine, however. We counted 32 total memory locations, which means that they are 3GB memory modules as that totals 96GB, the capacity of the card. 

Taking a closer look, we noticed some areas of poor contact along one edge of the inductors, where we noticed that half of the inductors are actually covered by a thermal pad. One of them only had about a third of the inductor covered. That’s not good as they really need to get the pad placement correct on GPUs that cost 8-11 thousand dollars.

The board also has lots of little tiny components, leading to a very dense PCB.  

Finally, we cleaned off the thermal paste off of the GPU, exposing the GB 202-870-A1 SKU. This is a 5090 die, just a fuller version of it.  

NVIDIA RTX 6000 Thermal Benchmarks

Thermal testing is up now. The RTX 5090 uses liquid metal with a 2-slot FE cooler that, genuinely, is one of the most impressive GPU coolers we’ve ever worked on. The prototype 4090 cooler we disassembled was more impressive in mechanical and cooling capabilities, but it was huge and impractical to manufacture. The 2-slot FE has done well in most cases, with the exception being memory cooling. We found that memory ran hot on our 5090 FE. That’s more of a concern with the PRO 6000 as well since it has so much memory.

Here’s a thermal test during a fixed render workload. The 5090 ran at around 72 degrees Celsius for the GPU core temperature in this test, which is overall excellent considering its 2-slot design. 

The VRAM for the 5090 ran at around 90 degrees Celsius, which was hotter than we’re comfortable with. Once accounting for installation into a case and potential long-term implications of pad dryout and dust, that’s high. The PRO 6000 ran its GPU core at a significantly hotter 82 degrees Celsius in this test, so about 10 degrees warmer than the 5090. The GPU memory was about the same, measuring 88 degrees via software.

Fan Response

None of that means anything without fan speeds.

Using only the auto settings, so following whatever VBIOS has programmed for the target GPU temperature, the RTX 5090 ran at 1550-1600 RPM for the average fan speed. The PRO 6000 ran at 1700. That’s a little faster with a much warmer core, but similar memory thermals.

Acoustics

Acoustics are next. For this testing, we’re using our hemi-anechoic chamber that we heavily invested in for improving our test quality. Our next major improvement will be upgrading the microphone, which is currently our bottleneck. We could hit a 6-8 dBA noise floor with a better mic, but we’re currently at 13.6 to 15.0 dBA for the floor. That’s still great though.

The RTX 5090 ran at about 32.5 dBA, with the PRO 6000 at the same noise level; however, the noise floor was slightly different between these two, so adjusting for that, the PRO 6000 would be about 1 dBA louder under the same conditions. This is hardly noticeable as a difference, if noticeable at all for most people.

Overall, the cooler design remains good when taking its size into consideration. The 5090 had higher noise levels in the higher frequency range, particularly 8000 to 10000 Hz. It also had a peak around 175 Hz, with the PRO peaking at 194 Hz and following a similar pattern, just adjusted right.

Generally speaking, this mostly follows the same trend. Our frequency cutoff is 150 Hz, so the higher result on the left can be ignored.

NVIDIA RTX 6000 Gaming Benchmarks

Gaming benchmarks are next. 

Dragon’s Dogma 2 - 4K

Dragon’s Dogma 2 is first. 

Tested at 4K, the RTX PRO 6000 landed at 140 FPS AVG, which has it ahead of the RTX 5090 by 5.8% for average framerate. Lows are not notably different. In other words, that’s about an extra $1,000 per 1% improvement, or about $780 per 1 FPS increase over the 5090 when calculating by the difference in cost for these two.

The lead in the RTX PRO 6000 is about 42% over the 4090 (watch our review), with the 5090 already ahead of the 4090 by about 34% here.

Dragon’s Dogma 2 - 1440p

At 1440p, the lead over the 5090 is expanded to 6.3%, ensuring we get our full gaming value out of the card. Low performance is not different outside of the usual scaling along with the improvement in average frametimes.

The RTX PRO 6000 hits 201 FPS AVG here, leading the 189 FPS of the 5090 and the 155 of the 4090. NVIDIA currently holds the entire top cluster of this chart, with AMD mostly focusing on the modern mid-range and Intel focused on the modern low-end.

Dragon’s Dogma 2 - 1080p

We next tested 1080p, which is what we all know buyers of this card really want to use it for. So-called AI use cases are obviously secondary compared to playing video games at 1080p.

The RTX PRO 6000 leads this extremely important and pivotal chart with a 226 FPS AVG, dropping to just a 5.3% lead. Honestly, there is something valuable that comes out of this chart: It’s marketing for AMD’s Ryzen 7 9800X3D, which is somehow keeping up with these GPUs enough that it can still produce distinguished results between the $2,000-$3,000 5090 and the $8,000 to $11,000 PRO 6000. That’s impressive. This is more amusing to us as it proves the longevity of our bench hardware more than anything else.

FFXIV 4K

Final Fantasy 14 is up now, tested first at 4K.

The RTX Pro 6000 ran at 208 FPS AVG, leading the 5090’s 184 FPS AVG result by about 13%. That’s a larger gain than we’ve seen in some of the other games, though obviously this is still a card intended for VRAM-intensive use cases and not gaming. That there’s still some power left beyond the 5090 is what’s more interesting, alongside the fact that the 9800X3D is so capable. Lows scale with the average and are not meaningfully better with the PRO 6000.

FFXIV 1440p

Tested at 1440p, we see a 10% improvement in framerate against the 5090, up at 348 FPS AVG for the 6000. Again, we’re mostly excited to see the 9800X3D continuing to scale and illustrating that there’s room left for another generation of testing on these benches. Lows remain in-step with the average.

FFXIV 1080p

At 1080p, the PRO 6000 again outperforms the 5090. This time, it’s reduced to about 4.3%. Let’s move on.

Starfield - 4K

Up next, we’re testing a card no one buys for gaming with a game no one plays. In Starfield at 4K, the PRO 6000 ran at 115 FPS AVG, leading the 5090 by about 7%. Lows are mostly within error. The value here is somewhere in the range of $700 to $1,000 per 1% improvement, depending on the price of the 5090.

Starfield - 1440p

At 1440p, the RTX PRO 6000 ran at 154 FPS AVG, leading the 147-148 FPS result of the 5090 marginally. The improvement over the 4090 is 17%, with the 5090 having previously led the 4090 by 12% in this test, as we’re becoming bound elsewhere.

The 1% lows for the 6000 are noteworthy here: At 89 FPS for the averaged 1% low, the PRO 6000 outperforms the average framerate of the 5070 and nearly matches the 4070 Ti with just the PRO’s 1% numbers alone. 

Resident Evil 4 - 4K

Resident Evil at 4K is up next. In this one, the RTX PRO 6000 ran at 219 FPS AVG, leading the 207 FPS AVG result of the 5090 by 6%. That’s consistent with most of the other tests so far. The lows, again, are not meaningfully different. The 4090’s 151 FPS AVG allows the 6000 a lead of 45%, or the 5090 a lead of 36.7%.

Resident Evil 4 - 1440p

At 1440p, the Blackwell workstation card ran at 371 FPS AVG, pushing the framerate measurably higher than the roughly 350 FPS AVG of the 5090 previously. We’re looking at about a 6.5% improvement, aligning with prior results again, and yet again showing just how good the 9800X3D is. 

Black Myth: Wukong - 4K

Black Myth: Wukong is up now, tested first at 4K and representing one of our heaviest non-RT workloads. The RTX PRO 6000 didn’t break the 100 FPS barrier, but it was the first card to exceed 90 -- not that this particularly matters to anyone when the 5090 was already in the mid 80s. The 6000 outperforms our 5090 FE result by 7.4%, above what we’ve seen on average thus far by about 1 percentage point. 92 FPS AVG is a good showing in this benchmark when at 4K. The next card in the stack is the 4090 at 67 FPS AVG, then the 5080 at 58 FPS AVG.

Black Myth: Wukong - 1440p

At 1440p, the PRO 6000 ran at 138 FPS AVG and led the 5090 marginally, but it was technically still a measurable improvement. Total uplift is 6.1%, aligning with Resident Evil 4 previously. Once again, the next card is the 4090. After this is the 5080, which isn’t that different from the 4080 Super (read our review), which isn’t that different from the 4080 (watch our review), which isn’t that different from the 5070 Ti. 

Black Myth: Wukong - 1080p

Black Myth: Wukong at 1080p is next. Selfishly, 1080p is interesting once again for probing at if there’s room in the CPU for more GPU. Turns out, there is: With the 9800X3D, the RTX PRO 6000 ran at 168 FPS AVG, improving on the 5090 by about 5%. That’s less than we saw at 1440p, which was less than we saw at 4K, so the card is gaining ground as resolution increases.

Dying Light 2 - 4K

Dying Light 2 is up next. At 4K, the PRO 6000 ran at 144 FPS AVG, which is a noteworthy improvement on the 126 FPS AVG of the 5090 FE. That’s a surprising 13.8%, which so far matches only one other game we’ve tested -- and that was Final Fantasy. Most of the other games are closer to 5-7%.

The 4090 is next at 91 FPS AVG, meaning the PRO 6000 is 59% ahead of the 4090.

Dying Light 2 - 1440p

At 1440p, there’s still scaling: The PRO 6000 is now at 246 FPS AVG, leading the 5090 by almost 13%. That’s a drop from what we saw at 4K, but still more than most other games.

Cyberpunk 2077: Phantom Liberty - 4K

Cyberpunk 2077: Phantom Liberty is one of our next most intensive games after Black Myth: Wukong. At 4K, the RTX PRO 6000 breaks through 100 FPS and hits 108, leading the 5090 by 13.5%. This is actually a pretty big lead and is another one of the games that has a larger average improvement. That’s still not worth $8,000 or $11,000 or whatever it is they’re selling these for, but it’s interesting to see any difference at all given the focus on so-called AI workloads.

Cyberpunk 2077: Phantom Liberty - 1440p

At 1440p, the PRO 6000 ran at nearly 200 FPS AVG, about 9.4% ahead of the 5090. This remains one of the larger gaps for 1440p, although again, it’s obviously not something anyone should remotely consider primarily for gaming. The 4090 is down at 135 FPS AVG, meaning the PRO 6000 leads it by 48%. 

Cyberpunk 2077: Phantom Liberty - 1080p

We’re leaving 1080p in because it’s the only place we see a bottleneck. Finally, we run into what seems like a CPU limit. The PRO 6000 and 5090 are at about the same level of performance, as is the 4090. Everything here is limited. That means our 1440p results were at least somewhat limited as well, with minimally the upper bound bouncing off of limits.

NVIDIA RTX 6000 Ray Tracing Benchmarks

Ray tracing benchmarks are next.

Ray Tracing - Black Myth: Wukong 4K

We’ll start with the heaviest one, which is Black Myth: Wukong at 4K with upscaling.

In this test, the RTX PRO 6000 ran at 92 FPS AVG, which was barely any different from the 88 FPS observed on the 5090. We’re under 5% of uplift here. It’ll be interesting to see if RT workloads don’t produce as much change as rasterization.

Ray Tracing - Black Myth: Wukong 1440p

At 1440p upscaled, the RTX PRO 6000 ran at 131 FPS AVG, leading the 5090 by just 3%. There’s just not much change in this test in general.

We’ll skip 1080p given the lack of improvement.

Ray Tracing - Cyberpunk 2077 (4K, RT Medium)

Cyberpunk at 4K with RT Medium is next. In this test, the RTX PRO 6000 ran at 65 FPS AVG, leading the 5090 by 11%. That’s roughly in-line with what we saw in Cyberpunk rasterized. This is without any upscaling, which may be helping the PRO card’s relative gain.

And finally, unlike what we saw in the 5060 and 5070 class cards with significant VRAM limitations harming low performance, the PRO doesn’t have any VRAM issues. This is clearly the solution. The solution is to buy $8K-$11K video cards.

As before, note that some cards with low VRAM, like the 5070 and 4070 Ti, have averages that look better than the reality (even though the averages are also not great).

Ray Tracing - Cyberpunk 2077 (4K, RT Ultra)

4K with RT Ultra is next. We almost never publish this test since it’s so intensive that it becomes somewhat useless, but it’s interesting here.

The RTX PRO 6000 ran at 56.6 FPS AVG, leading the 5090 by 6.4%. That’s a significant drop from what we saw at RT Medium, indicating that the increase in intensity for the RT workload is minimizing the benefit of other aspects of the PRO 6000’s improvements. It is getting overrun by the RT workload intensity. The lead over the 4090 is about 44% here for the 6000, with the 5090 leading the 4090 by 35%.

Note that the lower portion of this chart is unreliable for average framerate since the cards are all unplayable and stuttering. This can sometimes be due to exceeding VRAM limitations, such as on the RTX 5070. The average looks far better than the reality, even though the average is also unplayable.

Ray Tracing - Cyberpunk 2077 (1080p, RT Medium)

At 1080p with RT Medium, the PRO 6000 ran at 182 FPS AVG with lows paced proportionally. That has it ahead of the 5090 by about 12 FPS AVG, or about 7%.

Ray Tracing - Dragon’s Dogma 2 4K

Dragon’s Dogma 2 with ray tracing and without upscaling is next. Tested at 4K native, the RTX PRO 6000 at 121 FPS AVG leads the 5090 FE by about 7% in average framerate. This is starting to match a pattern, so we’ll let this be our last game test. Let’s move on to something else.

RTX 6000 AI Benchmark Charts:

Now, we'll get into some of our first ever LLM and machine learning benchmarks. As a disclaimer, we know enough to get some charts together to run some benchmarks where we feel like the controls of the test environment are good, but we do not yet know enough to have a full picture of capabilities outside of the test suite that we're using currently. 

We've tested with LM Studio, 3DMark testing, and ML Perf. We kind of settled on one set just to experiment for now. And this is our disclosure that we are experimenting with new testing, which means that the charts get the experimental chart label on top of them. And that's just so everyone's aware that these are not up to our full confidence standards of our normal benchmarks that we do all the time because we're still learning here.

LM Studio Testing

For our LM Studio testing, we recorded the response speeds from 8 different models that incrementally increase in size. 

We’ll get to larger models in a second.

Our first chart illustrates the AVG tokens/second results from the three smaller models – ranging from 8.5GB to around 19GB in size. We think these first three are more representative of like-for-like tests compared to some other models we’ll be going over today – because any of the GPUs on our chart can load these models without being limited by VRAM capacity.

In DeepSeek Llama 8B Distil, the RTX PRO 6000 is functionally tied for first place with the RTX 5090 FE – both with response speeds of roughly 81 tokens/second, improving on the 4090’s 62 tokens/second AVG by about 30%.

In 8-bit Phi-4, the PRO 6000 puts some distance between its own tokens/second AVG of 62 and the 5090's 51 tokens/second AVG, showing a 22% improvement. Compared to the 4090, the RTX PRO improves by about 43%.

In Qwen 2.5, the PRO 6000's 44 tokens/second AVG sees improvements of 25% over the 5090 and 37% over the 4090.

Text Generation

Moving to our next chart, we’ll really start to see performance gaps begin to widen as VRAM limitations take effect.

The RTX 6000 tops the chart in all five models.

In InternLM, the RTX PRO 6000's 50 tokens/second AVG improves upon the 5090's 40 by 25% and upon the standard 24GB 4090 by 319%.

In Mistral Small 26GB, the 4090 and 5090 see much sharper performance decreases than the PRO 6000. The workstation GPU achieves 42.4 tokens/second AVG, or 147% improved from the 5090’s 17 and 560% greater than the 4090’s AVG of 6.4.

Once we reach Gemma 3 27B, the workstation card really starts to separate itself from the lowly $2,000 gaming cards. In this model, the RTX PRO 6000 achieves a 29 tokens/second AVG, with the 5090 only reaching 5 and the 4090 only seeing 4 tokens/second, give or take some change. The comparisons remain largely the same in QwQ 32B.

In the final Llama 3.3 70b Q4_K_S model we tested, the RTX PRO 6000 achieved its greatest lead yet. The workstation card sees improvements of 928% over the 5090 and 1141% over the standard 4090. 

A more simplistic way to interpret this data is: At a certain point, the other cards run out of VRAM, while the workstation GPU doesn’t even use half of its VRAM in the largest Llama model we tested. You could expand beyond this as well, of course, but that’s the extent of our experimentation for now.

NVIDIA RTX 6000 Conclusion

That's it for the RTX Pro 6000 test. It's a review in a way but at the same time, we'd want to do a lot more tests in the ML category to really have a fully-fledged review. This piece has been an exciting experiment for us. 

The biggest discovery for us was seeing no liquid metal while we were doing the tear-down. That was probably known, but it’s not something that we had looked into. So that was interesting. We could see why NVIDIA did that for reliability reasons. We could also see it for liability reasons if the company is worried about liquid metal leaking out onto expensive servers. We kind of doubt this was done to save on costs, though maybe it was done for liability costs, but it just doesn't seem like NVIDIA is going to try and save a buck on this kind of card. 

We spoke to Wendell from Level1Techs and he noted that there's some apparent buggy behavior with Blackwell right now and so that's still getting updated and perhaps there's performance optimizations that could be done. On the side that we're familiar with, which is the gaming aspect, seeing 5 to 14% improvement is mostly useful not because you should ever in any way consider this card for gaming, because that would be an insane waste of money if that's all you would do with it, but more because it shows us that there's room left in the 9800X3D for scaling, which is pretty cool.

We compare Yeston’s “Waifu” RX 9070 XT vs PowerColor’s Red Devil equivalent in a series of tests that include acoustics, frequency, temps, gaming performance, and more

The Highlights

• Both RX 9070 XT cards cost $900
• The Red Devil 9070 XT needs to better tune its fan curves
• Both cards aren’t bad, but are overpriced at $900

Table of Contents

• AutoTOC

Intro

This is the most important review of all year. This is what makes or breaks this industry -- and even history: On one side, we have anime waifus from Atlantis bathed in the pleasant scent of the ocean vying for neckbeard attention, and on the other side, we have the Hellstone powering the Devil’s edgelord doomsayer RX 9070 XT. Only one will emerge victorious.

Editor's note: This was originally published on July 9, 2025 as a video. This content has been adapted to written format for this article and is unchanged from the original publication.

Credits

Test Lead, Host, Writing

Steve Burke

Testing

Mike Gaglione

Camera

Vitalii Makhnovets
Tim Phetdara

Animation, Editing

Andrew Coleman

Writing, Web Editing

Jimmy Thang

The Yeston RX 9070 XT Sakura Sugar Atlantis OC Edition weighs in at $900 and has its own YouTube channel devoted to Sakura, the angel of the ocean, including power move footage of soaking a video card in the very ocean sands whence it came. 

And when Yeston finally releases its much anticipated Husbando video card, named after the cool and aloof character “GAME ACE,” it’s GAME OVER for PowerColor’s Red Devil.

But Yeston isn’t alone in this battle of gamer marketing: PowerColor’s Hellhound says it has a “New Dawn” and “Power in Every Shade,” And... we don’t know what that means.

But it sounds scary -- almost as scary as the RED DEVIL, the natural counter to the WAIFU. The Red Devil says its Hellstone is “born from darkness, radiating power,” “forged in the fires of the underworld and inspired by the cradle of the rarest gems.” PowerColor “emerges as the unparalleled force in gaming, pushing performance to its absolute peak,” and “at its core lies the Hellstone, a breathtaking design element where vibrant RGB pulses with an otherworldly glow, embodying fiery power and precision.” 

Such a stone sounds too powerful to be in the hands of mere, filthy casuals. The breathtaking design element Hellstone is made of hundred-million-year-old materials dug from the depths of the earth. We’re told that it’s made of a rare, nearly unattainable polymeric matrix composite from hydrocarbon polymers, known only to mortal humans as... “plastic.”

This is the battle you all have been waiting for. It is the ultimate GPU sh*tpost -- second only to the one that NVIDIA just made by launching the so-called RTX 5050 (watch our coverage).

And if you’re wondering if this is just one gigantic meme with actual testing that required dozens of hours of work and hundreds of thousands of dollars worth of equipment all for this one meme, then yes. Yes it is. But it does have an important conclusion. And we are going to get to the bottom of who is better between the RED DEVIL VS. the WAIFU. Let’s battle. 

Testing

Despite the jokes, we do actually have testing in our review. We’re testing thermals, some power, acoustics, frequency, and a couple of games.  

The main reason we’re looking at games is because the 9070 XT Red Devil is running at a pretty high clock compared to other 9070 XTs so we wanted to look at that to see how much it actually mattered. 

The biggest issue with both cards is that they are currently $900, which is completely insane because MSRP is allegedly $600 for the 9070 XT, though that’s not really the case. Regardless, $900 is still pretty high. 

Both the cards we’re reviewing here have dual vBIOS and share a lot of common features. This means we’re going to focus most on aspects that are likely to be different. This includes thermals and acoustics. 

Thermals

The first test of prowess for these embattled icons is one of thermals: The Hellstone has a clear disadvantage here, mostly because it’s, uh, from hell, and the Waifu has an advantage with its affinity for oceans and Atlantis.

For this testing, our methodology involved testing the auto out-of-box settings in addition to testing with noise-normalized manual overrides to ensure an even battlefield. Waifus and Devils make different noises at different volumes, after all, and so it’s only fair to normalize them. 

The normalization required use of our hemi-anechoic chamber that we had custom-built exactly for purchases like this. You might think that spending $250,000 on a chamber to test the audio quality of meme video cards is insane, and you’d be right. That’s why we also test non-memes in it and use it every single benchmarking and technical reviews that we do for coolers and cases. To support our ability to buy test equipment like this and provide accurate data, head over to store.gamersnexus.net and grab one of our unique 3D Coaster Packs, like from our Debug pack. These have carefully placed and tested components to stabilize drinkware while providing a soft, rubberized material with PC theming. The debug code and power and reset switches on the red-and-black motherboard is one of our favorites, although the original coaster pack (also on the store) has the fan coaster that works surprisingly well and is also the most popular. Each pack comes with 4 coasters. Despite price increases over the years, we have maintained our original price. Head over to store.gamersnexus.net to grab these and support our in-depth testing and reviews -- even for memes.

Let’s get into the tests.

Thermals - 3DM 4K Auto VBIOS

The Waifu starts out with a strong showing when set to follow the default VBIOS profile and temperature targets with auto fan controls. The Atlantis-dweller ran at 57 degrees Celsius at steady state in a controlled ambient setting for GPU core on both the main and alternate VBIOS; in fact, as we showed in our Yeston Waifu review, the VBIOSes appear to do the same thing. It’s like they didn’t reprogram one of the fan curves or something, which may be a mistake.

But these temperatures are completely acceptable. They are maintained at 23.6 dBA and about 304W via GPU-Z board power monitoring.

The PowerColor Red Devil runs hotter: Its default OC VBIOS held a 60.7-degree result, with the silent VBIOS also seemingly the same. This has happened for multiple generations now with both Yeston and PowerColor. Maybe because the load hits steady state, they end up settling in the same place once under sustained load. We are restarting between each VBIOS change and even inspect the VBIOS profile name when switching to confirm that it does toggle. This process works to show differences in other devices, but at least here, we’re seeing the same across the profiles. That at least keeps it simple.

The Red Devil is running about 3-4 degrees warmer on the GPU and about 1-2 degrees warmer on the hotspot, with memory at a substantial climb of 8 degrees warmer. There’s a deficiency in memory cooling on the Red Devil by comparison, but an advantage in the hotspot-to-core delta, where Yeston’s delta is greater.

But there’s one other important factor: Out of the box, the PowerColor card is pulling 30W more via GPU-Z logging. This puts them at more equal thermal footing when considering the increased power consumption. The fan speed is another huge factor. So that was out of the box thermal performance, but we can start controlling these variables.

Thermals - 3DM 25 dBA

This test normalizes the GPUs for noise levels. We ran both at 25 dBA at 1-meter for this test.

The steady state 3DMark workload has the Waifu now 6 degrees warmer than the Red Devil in a searing blow to the Waifu, whose ocean and its scent-imbued plastic is evaporating before our very eyes. Boosting the Red Devil’s fan speeds to match the noise levels of the Waifu’s brings it down significantly, showing that it is the better card in like-for-like acoustic testing. Even with its smoldering Hellstone, the Red Devil has a 76-degree hotspot and delta of 26 degrees, improved from the Waifu’s 84-degree hotspot delta of 27.8-degree delta. This means the waifu is hotter here. Memory thermals are functionally tied between them, showing that PowerColor’s main area to improve is VRAM thermals. All of these numbers are acceptable, though, with the GPU numbers just straight-up good for both companies.

Notably, PowerColor is achieving these results with higher power draw. It lost when auto due to poor fan configuration, but pulls ahead when manually normalized.

Thermals - FurMark 25 dBA

In FurMark at 25 dBA, which loads the GPUs more heavily on the VRM, the GPU temperatures ended up comparable between these with the memory temperatures favoring the Waifu by about 2 degrees reduced from the Red Devil.

Acoustics

But this doesn’t mean much without acoustics. 

The most natural and first test of a Devil against a Waifu is the noise that they make in battle. For this reason, we threw these two GPUs into our battle arena: The hemi-anechoic chamber.

The Yeston RX 9070 XT Waifu edition ran at 1,130 RPM when under default VBIOS control at 304W in the workload that we ran. The card ran at 23.6 dBA at 1 meter when tested under the conditions we used for thermal testing. The frequency spectrum analysis has the highest spike around 164 Hz, with a swell around 500 Hz and a later bump and dip in the range of 1,000 Hz to 1,400 Hz.

The PowerColor Red Devil ran at 16.8 dBA, which is really quiet for what we would imagine of hell. It’s too close to our floor to be all that useful. The Devil is substantially quieter out of box, mostly because it just isn’t utilizing its fans as well as we think it should be -- that’s why it ran warmer in the auto testing. We did some in-depth investigating and realized it’s because the Hellstone converts the Red Devil’s angry howlings into the power needed to sustain its own power consumption, destroying the sound waves in the process. Fans ran at 910 RPM rather than 1,130 RPM, which means it’s quiet, but we think also not properly utilized. PowerColor could find more of a balance here, although maybe they’re trying to remain true to the fiery marketing.

The frequency spectrum has the Devil’s peak at around 320 Hz, with an elevated area around 420 Hz. Overall, it’s a lot quieter than the Waifu, with the high-end of the frequency scale similar.

Acoustics - Noise-Normalized

When normalized to roughly the same noise levels, leaving the Yeston card on the chart, the Red Devil’s new frequency spectrum has the Devil’s peak at about 200 Hz. There’s a slight wave up around 400-500 Hz and another around the same place as the Yeston card, at 900 Hz, then a dip, then up again at around 1,200 Hz.

Frequency

GPU frequency is the next test in this battle of the memes. The Yeston GPU ran at about 2,890-2,910 MHz in this workload when running a fixed frame render. The Red Devil ran notably higher, up at 2,970-3,000 MHz instead. That’s around a 60-100 MHz bump in most instances, so the extra power is going somewhere, and it’s sustaining that frequency.

Plotting the Sapphire Pulse as a frame of reference, it holds about 2,900-2,925 MHz at steady state, putting it slightly faster than the Yeston card. The Waifu’s desired frequency just happens to be lower than that of the Pulse and Red Devil.

Gaming Performance

We’ll do a quick look at gaming performance to see how much that extra frequency does or doesn’t matter. This will be a very slimmed-down version of our tests since we only need a few head-to-head results. These were tested on the same drivers, so for purposes of comparing the 9070 XTs to each other, that’s all we need.

FFXIV - 4K

Final Fantasy is first. We chose this to start because the ratio of Waifu-to-non-Waifus in Final Fantasy is extremely disproportionate, and so Yeston should have a natural advantage.

The Red Devil ran at 69.6 FPS AVG in Final Fantasy 14 Dawntrail at 4K, leading the Yeston Atlantis Waifu card by just over 1 FPS AVG. You’d never notice this, but on a technicality, the Red Devil is about 1.9% higher average framerate. The Pulse is about the same as the Atlantis card.

Resident Evil 4 - 1440p

Naturally, because we biased the Final Fantasy test toward the Waifu by nature of the... waifus all over the game, we had to bias the next one toward the Red Devil. Resident Evil 4 is up now, giving PowerColor a distinct and hellish advantage. 

The RX 9070 XT Red Devil draws upon the very evils which it is rendering with its Hellstone to run at 198 FPS AVG, about a 3 FPS lead over the Waifu. That’s about 1.7% ahead. The Waifu outruns the Pulse, though, and you don’t need to be the fastest when you’re running from zombies -- you just need to not be the slowest.

Resident Evil 4 - 4K

The only thing worse than zombies at 1440p is zombies at 4K, so that’s what we’re rendering now. In Resident Evil 4 at 4K, the Red Devil held a 106 FPS AVG, leading the Waifu by about 2 FPS AVG. That’s around 2%.

All of these numbers are expected. We did technically run 4 other games, but they all show the same thing: Typically, we’re in the range of 1-3% improved on the PowerColor card out of the box.

Conclusion

The price for both cards is the same, and they’re both crazy. The good news is that because both are $900, they’re easy to compare.

We have to give both cards credit. Neither are bad. We’ve already reviewed the Yeston RX 9070 XT Sakura Sugar Atlantis OC Edition, in which we also tore down the card. We were fairly positive on it. There were things it could have done better. For example, it doesn’t have as good of a thermal solution as we’ve seen on some of their past cards with this design approach, but it was still fine in pretty much every metric. 

For the Red Devil, the biggest place that card is lacking is in underutilizing the ability to tune its fan curves more. This is something PowerColor seems to consistently have problems with in the cards we test from them. The company could utilize it more to better balance the thermals and acoustics more reasonably.

Overall, though, both cards are pretty close to each other. The Red Devil, in noise-normalized situations, tends to be the better cooler, but the Waifu card, out of the box, tends to be the better cooler. It depends on how you’re going to run them. The Red Devil also has a higher clock out of the box and it actually runs faster. Now, in gaming performance, this difference doesn’t really manifest itself in a big way. 

We did play around with overclocking on both cards. We got a little higher clocks out of the Red Devil. That card also has a higher power budget.

Both cards are too expensive, however. At least both cards do something to try and justify their price.

We’ve compiled a comprehensive timeline of the GPU bans, GPU smuggling, and export controls that impact NVIDIA, AMD, and Intel

The Highlights

• The US blocks exports of advanced GPUs to China to protect national security
• NVIDIA GPUs are highly sought after in China for AI processing
• Our timeline chronicles the US export controls, NVIDIA's responses, and reports of GPU smuggling

This is a comprehensive timeline of the GPU bans, smuggling, and export controls on NVIDIA, AMD, Intel, and other high-tech semiconductor products. We are publishing this as part of our stretch goals for Black Market AI GPU — a viewer-funded film made possible through support on our store, including our new “Blind Eye” T-shirt.
The below timeline accompanies our Black Market AI GPU investigation, our biggest project yet. We spent three weeks in Asia to uncover this story, including two weeks in China and one in Taiwan. We found smugglers, middlemen, and users of so-called “AI” GPUs that the United States government has banned for sale into China.

Credits

Host, Writing, Lead Editing

Steve Burke

Editing

Vitalii Makhnovets
Tim Phetdara

Editing, Graphics

Andrew Coleman

Camera

Tannen Williams

Research and Writing

Ben Benson

We are providing this timeline for free and without third-party ads for our viewers and readers. As this situation has changed frequently and now spans multiple US administrations, we may have missed a few events. However, we believe we have compiled all the major changes – especially since the start of 2025 – that are directly relevant to the story. 

We have attempted to present it as neutrally as feasible and from a place of reporting. We’ve included links to a variety of media and government sources that we believe are appropriate for establishing the timeline of events. We have included statements from NVIDIA in many cases. 

This was a huge team effort at GN and required a massive investment in travel, writing, research, and editing to complete. If you find this information valuable, we ask that you please support us directly by backing our NVIDIA AI GPU Black Market project, buying something from our GN store, or signing up for our Patreon. Thank you.

Timeline

Note on sources: Our intent is to cite primary sources, including government documents, and a variety of secondary sources. In some cases, we link only to secondary news stories. This can occur when we include articles from credible media reports but do not have primary documents to cite. 

2018 GPU Export Controls

August 2018

August 13: The US government created the National Security Commission on Artificial Intelligence (NSCAI) as part of the John S. McCain National Defense Authorization Act for Fiscal Year 2019. The NSCAI had 15 commissioners who were nominated by Congress and the Executive Branch. The NSCAI was tasked with investigating how the United States should compete in AI in the modern age and recommending actions for Congress and the executive branch.

In the words of the original document, the commissioners “shall consider the methods and means necessary to advance the development of artificial intelligence, machine learning, and associated technologies by the United States to comprehensively address the national security and defense needs of the United States.”

• Source:
  • National Defense Bill (NSCAI section starts on page 1963)

2019 GPU Export Controls

May 2019

May 15: Citing national security risks, the US government added Huawei to its Entity List and restricted sales of Huawei’s equipment into the United States. 

• Sources:
  • Federal Register
  • Executive Order
  • Politico: Trump signs order setting stage to ban Huawei from U.S.
  • Reuters: Trump administration hits China's Huawei with one-two punch
  • Reuters: China's Huawei, 70 affiliates placed on U.S. trade blacklist

2020 GPU Export Controls

May 2020

May 19: The United States restricted semiconductor designs, chipsets, and technologies to Huawei and its foreign affiliates. 

• Sources: 
  • Department of Commerce Press Release
  • Federal Register
  • NPR: The Latest U.S. Blow To China's Huawei Could Knock Out Its Global 5G Plans

2021 GPU Export Controls

March 2021

The National Security Commission on Artificial Intelligence (NSCAI) released its final report. The report provided recommendations to “advance the development of artificial intelligence, machine learning, and associated technologies to comprehensively address the national security and defense needs of the United States.”

As part of the report (page 216), the NSCAI recommended the US government and its allies “utilize targeted export controls on high-end semiconductor manufacturing equipment… to protect existing technical advantages and slow the advancement of China’s semiconductor industry.” 

Further on (page 228), the report said, “Looking across the AI stack, the hardware component of the AI stack contains the most viable targets for traditional export controls.” The report (page 231) focused on semiconductor manufacturing equipment for export control rules: “The primary U.S. export control target to constrain competitors’ AI capabilities should be sophisticated semiconductor manufacturing equipment (SME) necessary to manufacture high-end chips.” 

The report mentioned export controls for GPUs (page 500) as a way “to prevent the use of

high-end U.S. AI chips in human rights violations.”

• Source: 
  • NSCAI Final Report – The PDF can be downloaded from here.

April 2021

NSCAI Commissioner Christopher Darby spoke at NVIDIA GTC about the NSCAI’s report to Congress.

• Source: 
  • GTC

October 2021

October 1: The NSCAI officially ended on October 1, 2021. 

• Source: 
  • Executive Government News

2022 GPU Export Controls

August 2022

August 31: NVIDIA filed a Form 8-K with the SEC to inform investors that the US government had immediately blocked exports of its A100 and H100 chips to China, including Hong Kong. The export controls included DGX or other systems that incorporate an A100, H100, or A100X. In the financial documents, NVIDIA said the US government informed it of the export restrictions on August 26, 2022. NVIDIA stated that its third-quarter results included up to $400 million in expected sales to China that were now uncertain due to the export restrictions.

• Sources: 
  • NVIDIA Form 8-K
  • NVIDIA Form 10-Q
  • SEC Filing

September 2022

September 1: NVIDIA filed a new Form 8-K to let customers know that the US government had offered some exemptions for certain chip exports: 

“The U.S. government has authorized exports, reexports, and in-country transfers needed to continue NVIDIA Corporation’s, or the Company’s, development of H100 integrated circuits after the Company filed its Current Report on Form 8-K with the U.S. Securities and Exchange Commission on August 31, 2022. The authorization also allows the Company to perform exports needed to provide support for U.S. customers of A100 through March 1, 2023. Additionally, the U.S. government authorized A100 and H100 order fulfillment and logistics through the Company’s Hong Kong facility through September 1, 2023.”

• Source: 
  • NVIDIA Form 8-K

Following NVIDIA’s SEC filing, media outlets reported the US government ordered NVIDIA to stop selling advanced AI chips to China. 

• NVIDIA statement: 

• “We are working with our customers in China to satisfy their planned or future purchases with alternative products and may seek licenses where replacements aren’t sufficient. The only current products that the new licensing requirement applies to are A100, H100 and systems such as DGX that include them.”– NVIDIA to CNBC

• Sources:
  • CNBC: NVIDIA stock falls after U.S. government restricts chip sales to China
  • The New York Times: U.S. Restricts Sales of Sophisticated Chips to China and Russia
  • Reuters: U.S. officials order NVIDIA to halt sales of top AI chips to China
  • CNN: US orders NVIDIA and AMD to stop selling AI chips to China 
  • Associated Press: China demands US drop tech export curbs after NVIDIA warning

October 2022

October 7: The Department of Commerce’s Bureau of Industry and Security (BIS) implemented a series of export controls to “protect US national security and foreign policy interests.” The new export controls would hinder China’s ability to build high-end semiconductors and purchase advanced chips from the US, including for development of and maintaining supercomputers. 

• Sources: 
  • Department of Commerce Release
  • Federal Register (amended on October 13)

In a briefing with reporters, the US government said the new regulations formalized the guidance previously sent to NVIDIA. The Guardian reported: 

“The new regulations will also severely restrict export of US equipment to Chinese memory chip makers and formalize letters sent to NVIDIA Corp and Advanced Micro Devices Inc (AMD) restricting shipments to China of chips used in supercomputing systems that nations around the world rely on to develop nuclear weapons and other military technologies.”

• Sources: 
  • The Guardian: Biden administration imposes sweeping tech restrictions on China
  • The New York Times: Biden Administration Clamps Down on China’s Access to Chip Technology

November 2022 

November 7: Reuters reported that NVIDIA had created a new AI chip called the A800 GPU for the China market. The A800 would be compliant with US export controls.

• NVIDIA statement:

• “The NVIDIA A800 GPU, which went into production in Q3, is another alternative product to the NVIDIA A100 GPU for customers in China. The A800 meets the US government’s clear test for reduced export control and cannot be programmed to exceed it.” - NVIDIA to Reuters

• Sources:
  • Reuters: Exclusive: NVIDIA offers new advanced chip for China that meets U.S. export controls
  • Tom’s Hardware: NVIDIA Creates New Supercomputer Chip For Chinese Market
  • The Verge: NVIDIA’s selling a nerfed GPU in China to get around export restrictions
  • TechCrunch: NVIDIA touts a slower chip for China to avoid US ban

2023 GPU Export Controls

March 2023

March 21: Reuters reported that NVIDIA had modified the H100 to be compliant with export rules to China.

• NVIDIA’s statements to Reuters:

• “On Tuesday, the company said it has similarly developed a China-export version of its H100 chip. The new chip, called the H800, is being used by the cloud computing units of Chinese technology firms such as Alibaba Group Holding Ltd, Baidu Inc and Tencent Holdings Ltd, a company spokesperson said.” [...]

“The NVIDIA spokesperson declined to say how the China-focused H800 differs from the H100, except that ‘our 800 series products are fully compliant with export control regulations.’” 

• Sources: 
  • Reuters: NVIDIA tweaks flagship H100 chip for export to China as H800
  • Data Center Dynamics: NVIDIA creates pared back H100 GPU for export to China, called H800

June 2023

June 27: The Wall Street Journal reported that the US government is considering expanding export controls for GPUs and AI chips to China. The US Department of Commerce did not comment to the Wall Street Journal. 

• Sources:
  • The Wall Street Journal: U.S. Considers New Curbs on AI Chip Exports to China
  • MarketWatch: White House says it’s focused on being at front end of supply chain for chips, won’t comment on report of possible new ban on exporting AI chips to China
  • Reuters: US mulls new export restriction on computing power in AI chips
  • TechCrunch: China’s AI firms might further lose chip access in new US ban

October 2023

October 17: The US Department of Commerce updated its export compliance for advanced semiconductors and semiconductor manufacturing equipment. The government said: 

“Today’s rules reinforce the October 7, 2022, controls to restrict the PRC’s ability to both purchase and manufacture certain high-end chips critical for military advantage. These updates are necessary to maintain the effectiveness of these controls, close loopholes, and ensure they remain durable.” 

The U.S. government has removed interconnect speed as a criterion for identifying restricted chips. Instead, it will now focus on processor performance and performance density. In a statement, the government said:

“A performance density parameter prevents the workaround of simply purchasing a larger number of smaller datacenter AI chips which, if combined, would be equally powerful as restricted chips.”

• Sources: 
  • Department of Commerce Release
  • Federal Register
  • Bureau of Industry and Security Document on Performance Density (page  21-22)
  • The Register: Biden has brought the ban hammer down on US export of AI chips to China
  • Center for Strategic & Internal Studies (posted on Oct. 18)

As part of the announcement, the administration told reporters the new restrictions affect NVIDIA’s A800 and H800 chips. A few days prior, Reuters reported that the administration would soon announce new export rules. 

• Sources:
  • Reuters: Exclusive: US tackles loopholes in curbs on AI chip exports to China
  • CNBC: U.S. curbs export of more AI chips, including NVIDIA H800, to China
  • Reuters: NVIDIA details advanced AI chips blocked by new export controls
  • Tom’s Hardware: US Prohibits Exports of NVIDIA’s A800 and H800 to China, Blacklists Chinese GPU Developers
  • The Verge: NVIDIA’s H800 AI chip for China is blocked by new export rules

October 23: NVIDIA filed a Form 8-K with the SEC that said the new export rules impact its A100, A800, H100, H800 and L40S chips. NVIDIA said it “does not anticipate that the accelerated timing of the licensing requirements will have a near-term meaningful

impact on its financial results.” 

• Sources: 
  • NVIDIA Form 8-K
  • BBC: US orders immediate halt to some AI chip exports to China, NVIDIA says
  • Tom’s Hardware: US Govt Speeds Up Export Restrictions for NVIDIA’s GPUs

December 2023

December 6: In a meeting with reporters in Singapore, NVIDIA said that it was working on new chips that comply with the government’s rules. 

• NVIDIA statement:

• “NVIDIA has been working very closely with the U.S. government to create products that comply with its regulations. Our plan now is to continue to work with the government to come up with a new set of products that comply with the new regulations that have certain limits.” – NVIDIA CEO Jensen Huang, as reported in Reuters

• Sources:
• Reuters: NVIDIA working closely with US to ensure new chips for China are compliant with curbs
  • Reposted in CNBC

December 28: NVIDIA released a new version of RTX 4090 for the China market. The new chip, called the GeForce RTX 4090D, would be compliant with US export control restrictions. 

• NVIDIA statements: 

• “The GeForce RTX 4090 D has been designed to fully comply with U.S. government export controls. While developing this product, we extensively engaged with the U.S. government.” - NVIDIA to Reuters

• “In 4K gaming with ray tracing and deep-learning super sampling (DLSS), the GeForce RTX 4090D is about five percent slower than the GeForce RTX 4090 and it operates like every other GeForce GPU, which can be overclocked by end users.” – NVIDIA to The Register

• Sources: 
  • Reuters: NVIDIA launches new gaming chip for China to comply with US export controls
  • The Register: NVIDIA slowed RTX 4090 GPU by 11 percent, to make it 100 percent legal for export to China
  • The Verge: NVIDIA is releasing a slower RTX 4090 in China to comply with US restrictions

2024 GPU Export Controls

February 2024

February 1: Reuters reported that NVIDIA had prepared new GPUs for China, including the H20. Several sources told Reuters that the new offerings are less powerful than similar chips from Huawei. 

• Sources:
  • Reuters: Exclusive: NVIDIA’s new China-focused AI chip set to be sold at similar price to Huawei product
  • Tom’s Hardware: New NVIDIA AI GPUs designed to get around U.S. export bans come to China — H20, L20, and L2 to fill void left by restricted models

July 2024

July 22: Reuters reported that NVIDIA is creating a new GPU for the China market based on its Blackwell chips. Sources told Reuters that the chip would be a version of the Blackwell B200. NVIDIA did not publicly disclose the specifications. 

• Sources:
  • Reuters: Exclusive: NVIDIA preparing version of new flagship AI chip for Chinese market
  • Tom’s Hardware: NVIDIA preparing a China-focused variant of its B200 Blackwell AI GPU to comply with US export regulations
  • HPCWire: NVIDIA Prepares New AI Chip for China Amid Ongoing US Export Controls

December 2024

December 2: The US government expanded rules that limit the export of high memory bandwidth (HBM) and advanced semiconductor equipment. 

The Center for Strategic & International Studies explained the new rules on HBM: 

“The December 2024 controls change that by adopting for the first time country-wide restrictions on the export of advanced HBM to China as well as an end-use and end-user controls on the sale of even less advanced versions of HBM. The goal of these controls is, unsurprisingly, to degrade China’s AI industry.” [...]

“Modern AI chips not only require a lot of memory capacity but also an extraordinary amount of memory bandwidth. Bandwidth refers to the amount of data a computer’s memory can transfer to the processor (or other components) in a given amount of time. With low-bandwidth memory, the processing power of the AI chip often sits around doing nothing while it waits for the necessary data to be retrieved from (or stored in) memory and brought to the processor’s computing resources.”

• Sources:
  • Department of Commerce Document
  • Federal Register 
  • Government Presentation
  • Center for Strategic & International Studies: Understanding the Biden Administration’s Updated Export Controls

2025 GPU Export Controls

January 2025

January 13: The US government tightened its export controls by introducing national chip caps for many countries, except for 18 allies. The new restrictions would be called the AI Diffusion Rule. The rule would go into effect in May 2025.  

• NVIDIA statement:

• “It makes no sense for the Biden White House to control everyday datacenter computers and technology that is already in gaming PCs worldwide, disguised as an anti-China move. The extreme ‘country cap’ policy will affect mainstream computers in countries around the world, doing nothing to promote national security but rather pushing the world to alternative technologies. AI is mainstream computing – ubiquitous and essential as electricity. This last-minute Biden Administration policy would be a legacy that will be criticized by U.S. industry and the global community. We would encourage President Biden to not preempt incoming President Trump by enacting a policy that will only harm the U.S. economy, set America back, and play into the hands of U.S. adversaries.” – Ned Finkle, Vice President of Government Affairs, NVIDIA, to Bloomberg (Twitter link)

• Sources: 
  • US Government Fact Sheet
  • Federal Register
  • Federal Register
  • Bloomberg: Biden to Further Limit NVIDIA AI Chip Exports in Final Push
  • Tom’s Hardware: NVIDIA and SIA fire back at US gov's new export restrictions on AI GPUs to China

February 2025

February 26: NVIDIA filed its 10-K annual report with the SEC. In the 10-K, NVIDIA revealed that Singapore was the second-largest geographical source of revenue in 2024, behind the United States. Taiwan was third, and China was fourth.

Within the report, NVIDIA said: 

“Singapore represented 18% of fiscal year 2025 total revenue based upon customer billing location. Customers use Singapore to centralize invoicing while our products are almost always shipped elsewhere. Shipments to Singapore were less than 2% of fiscal year 2025 total revenue.”

• Source:
  • NVIDIA 10-K

February 27: Speculation began about AI GPUs being smuggled from Singapore to China. In late February, authorities in Singapore arrested three people for fraud involving servers that may contain AI GPUs. Singapore’s government granted the three people bail a few weeks later.

NVIDIA declined to comment to CNBC. 

• Sources:
  • ChannelNewsAsia: 3 men charged with fraud, cases linked to alleged movement of Nvidia chips
  • CNBC: NVIDIA’s unofficial exports to China face scrutiny after arrest of silicon smugglers in Singapore
  • Tom’s Hardware: Singapore police bust major ring smuggling NVIDIA GPUs to China-based DeepSeek: Report
  • TechCrunch: Singapore grants bail for NVIDIA chip smugglers in alleged $390M fraud

April 2025

April 9: NPR reported that the US government would not add export controls for the H20 chip after NVIDIA CEO Jensen Huang attended a dinner at Mar-A-Lago. The dinner reportedly cost $1 million per head. The outlet said it was unclear whether Jensen Huang met with US President Trump directly. 

NVIDIA declined to comment to NPR.

• Sources:
  • NPR: Trump administration backs off NVIDIA H20 chip crackdown after Mar-a-Lago dinner

April 15: In a SEC filing, NVIDIA said the US government sent the company new export rules on April 9. According to NVIDIA, the H20 and all chips with the H20’s memory bandwidth or interconnect bandwidth will now need licenses to export to China. NVIDIA said the new rules would cost the company $5.5 billion in charges due to current H20 chip inventory and prior sales. NVIDIA declined to comment further to the BBC. 

• Sources: 
  • NVIDIA Form 8-K
  • BBC: NVIDIA shares plunge amid $5.5bn hit over export rules to China
  • NPR: NVIDIA discloses that U.S. will limit sales of advanced chips to China after all
  • Reuters: US issues export licensing requirements for NVIDIA, AMD chips to China

April 16: The US government released an investigative report on DeepSeek and requested information from NVIDIA about its AI GPUs. Through a letter sent to NVIDIA CEO Jensen Huang, the US government asked NVIDIA for a list of its customers in China and many countries in Asia, including Singapore. The government requested all communication between NVIDIA and DeepSeek. 

• Sources:
  • US government press release
  • DeepSeek report
  • Government letter to Jensen Huang

The US Department of Commerce confirmed that it has issued new export control rules for AI chips. The Commerce Department provided a statement to The Wall Street Journal: 

“The Commerce Department is issuing new export licensing requirements on the NVIDIA H20, AMD MI308, and their equivalents.”

• Sources:
  • The Wall Street Journal: U.S. Confirms New Export Curbs on NVIDIA and AMD Chips

April 28: The Wall Street Journal reported that Huawei is expected to release its new AI chip, the Ascend 910D, soon. According to the Wall Street Journal’s sources, Huawei expects the Ascend 910D to be about as powerful as an NVIDIA H100. 

• Sources:
  • The Wall Street Journal: China’s Huawei Develops New AI Chip, Seeking to Match NVIDIA 
  • NetworkWorld: Huawei steps up AI chip race with Ascend 910D, targeting NVIDIA’s high ground

April 30: Anthropic, an AI startup backed by Amazon, called on the US government to increase export control restrictions to China. As part of a blog post, Anthropic said the government needs to improve its export enforcement to reduce smuggling. The company cited examples of chips being smuggled with “prosthetic baby bumps” and “live lobsters.”

In a response, NVIDIA said: 

“American firms should focus on innovation and rise to the challenge, rather than tell tall tales that large, heavy, and sensitive electronics are somehow smuggled in ‘baby bumps’ or ‘alongside live lobsters.’” – NVIDIA to CNBC

• Sources:
  • Anthropic blog post
  • CNBC: NVIDIA says Anthropic is telling ‘tall tales’ in its defense of U.S. AI chip restrictions on China
  • Tom’s Hardware: Despite NVIDIA claims, Chinese smugglers have used live lobsters and fake baby bumps to traffic chips
  • Hong Kong Customs release
  • China Customs release 

May 2025

May 1: Jensen Huang spoke with the House Foreign Affairs Committee to discuss domestic manufacturing and the importance of AI. NVIDIA posted the remarks online.

• Sources:
  • NVIDIA Newsroom Twitter Post
  • Tom’s Hardware: NVIDIA warns U.S. AI hardware export rules could backfire, empowering Huawei to define global standards

May 7: Following a report in Bloomberg, the US Department of Commerce confirmed that it will not implement the AI Diffusion Rule that was created during the prior administration. The rule would have gone into effect on May 15, 2025. 

The Department of Commerce released a statement to CNBC: 

“The Biden AI rule is overly complex, overly bureaucratic, and would stymie American innovation. We will be replacing it with a much simpler rule that unleashes American innovation and ensures American AI dominance.” 

NVIDIA released a statement: 

“We welcome the Administration’s leadership and new direction on AI policy. With the AI Diffusion Rule revoked, America will have a once-in-a-generation opportunity to lead the next industrial revolution and create high-paying U.S. jobs, build new U.S.-supplied infrastructure, and alleviate the trade deficit.”

• Sources: 
  • Bloomberg: Trump to Rescind Global Chip Curbs, Prep New AI Restrictions
  • Tom’s Hardware: NVIDIA celebrates dumping of Biden-era AI chip export rules — simpler new policy promised
  • CNBC: Trump administration set to end Biden’s U.S. chip export restrictions
  • NVIDIA Twitter Account

May 9: Reuters reported that NVIDIA is preparing a cut down version of the H20 for the Chinese market. Reuters sources said the chip would be ready in July. NVIDIA declined to comment. 

• Sources: 
  • Reuters: Exclusive: NVIDIA modifies H20 chip for China to overcome US export controls, sources say
  • Tom’s Hardware: NVIDIA readies cut-down HGX H20 GPU for China to comply with export control rules

May 13: The US government formally rescinded the previous administration’s AI Diffusion Rule, which was announced in January 2025. The government also announced actions to strengthen export controls for AI chips, including restrictions on using several Huawei Ascend chips

NVIDIA declined to comment to CNBC on the new export restrictions. 

• Sources:
  • US government press release 
  • BIS policy statement 
  • US government guidance on using Huawei Ascend chips
  • The Register: Trump ends Biden-era dream to cap US AI chip exports
  • Bloomberg: US Warns That Using Huawei AI Chip ‘Anywhere’ Breaks Its Rules
  • Tom’s Hardware: U.S. issues worldwide crackdown on using Huawei Ascend chips, says it violates export controls
  • CNBC: Trump administration’s next wave of China AI chip export rules are yet another obstacle for NVIDIA 

May 15: A bipartisan group of legislators introduced the Chip Security Act that is intended to stop smuggling of high-end AI chips. 

The Hill summarized the proposed legislation: “The legislation, titled the Chips Security Act, would require companies to ensure the location-verification abilities of their high-end AI chips and to report when a product has been diverted or changed location. It follows recent reports of increased smuggling of chips, including those made by NVIDIA, into China despite tight export controls.”

NVIDIA declined to comment to The Register.

• Sources: 
  • Chip Security Act text
  • The Hill: Bipartisan House lawmakers propose bill to ‘stop smuggling’ of AI chips
  • Reuters: U.S. lawmakers introduce bill to address AI chip smuggling
  • The Register: Plan to keep advanced chips from China with tracking tech gains support in Congress

May 16: The Financial Times reported that NVIDIA intends to create a research and design center in Shanghai. 

• NVIDIA statement: 

• “We are not sending any GPU designs to China to be modified to comply with export controls.” - NVIDIA to CNBC 

• Sources: 
  • Financial Times: NVIDIA plans Shanghai research centre in new commitment to China
  • CNBC: NVIDIA says it is not sending GPU designs to China after reports of new Shanghai operation

May 19: NVIDIA CEO Jensen Huang told Bloomberg in a TV interview that he didn’t see any “evidence” of any AI chip diversion. Tom’s Hardware summarized Jensen Huang’s quote: 

“Governments understand that diversion is not allowed, and there’s no evidence of any AI chip diversion — recognize our data center GPUs are massive; these are massive systems. The Grace Blackwell system is nearly two tons, and so you’re not going to be shipping — you’re not going to be putting that in your pocket or your backpack anytime soon. And so, these systems are fairly easy to keep track of... but the important thing is that the countries and the companies that we sell to recognize that diversion is not allowed, and everybody would like to continue to buy NVIDIA technology, and so they very well monitor themselves very carefully and they’re quite careful about that.”

• Sources:
  • Bloomberg: NVIDIA CEO Says ‘No Evidence of Any AI Chip Diversion’
  • Tom’s Hardware: NVIDIA CEO Jensen Huang says ‘There’s no evidence of any AI chip diversion’

May 21: At Computex 2025, NVIDIA CEO Jensen Huang called the US export controls a “failure.” He said that NVIDIA’s market share in China has dropped from 95% to 50% due to the restrictions. The Guardian quoted Jensen Huang as saying: 

“The local companies are very, very talented and very determined, and the export control gave them the spirit, the energy and the government support to accelerate their development.” [...]

“I think, all in all, the export control was a failure.” [...]

“China has a vibrant technology ecosystem, and it’s very important to realise that China has 50% of the world’s AI researchers, and China is incredibly good at software.”

• Sources: 
  • CNBC: Jensen Huang says U.S. chip restrictions have cut NVIDIA’s China market share nearly in half
  • The Guardian: US chip export controls are a ‘failure’ because they spur Chinese development, NVIDIA boss says
  • The New York Times: NVIDIA’s Chief Says U.S. Chip Controls on China Have Backfired

May 27: Reuters reported that NVIDIA plans to launch a new, cheaper Blackwell-based GPU for the China market to comply with US export rules. 

• NVIDIA statement: 

• "Until we settle on a new product design and receive approval from the U.S. government, we are effectively foreclosed from China's $50 billion data center market." - NVIDIA to Reuters

• Sources: 
  • Reuters: Exclusive: NVIDIA to launch cheaper Blackwell AI chip for China after US export curbs, sources say
  • SiliconANGLE: Report: NVIDIA racing to develop new, scaled-down Blackwell GPUs for China

May 28: During NVIDIA’s quarterly earnings, CEO Jensen Huang said the company was writing off unsold H20 inventory due to export controls. VentureBeat posted Jensen Huang’s quote from earnings: 

“Let me share my perspective on some topics we’re frequently asked on export control. China is one of the world’s largest AI markets and a springboard to global success with half of the world’s AI researchers based there. The platform that wins China is positioned to lead globally today. However, the $50 billion China market is effectively closed to U.S. industry. The H20 export ban ended our Hopper data center business in China. We cannot produce Hopper further to comply. As a result, we are taking a multibillion-dollar write-off on inventory that cannot be sold or repurposed. We are exploring limited ways to compete, but hopper is no longer an option.”

• Sources:
  • VentureBeat: NVIDIA CEO takes a shot at U.S. policy cutting off AI chip sales to China
  • PC Gamer: NVIDIA’s Hopper GPUs are now dead to the Chinese market after export controls that made the company take a 'multibillion-dollar write-off'

June 2025

June 12: NVIDIA’s CEO Jensen Huang told CNN the company will no longer include sales and revenue from China in its forecasts. In a response to a question from CNN about whether the US government would lift its export controls, Jensen Huang said: 

“I’m not counting on it but, if it happens, then it will be a great bonus. I’ve told all of our investors and shareholders that, going forward, our forecasts will not include the China market.” 

• Sources:
  • CNN: NVIDIA will stop including China in its forecasts amid US chip export controls, CEO says

June 18: Several media reported on rumors about NVIDIA preparing to launch a “RTX 5090 DD” for the China market. The new card would allegedly reduce the memory specifications compared to the RTX 5090D. 

• Sources: 
  • Tom’s Hardware: NVIDIA planning new RTX 5090 'DD' variant for China — 24GB card with tweaked GPU latest attempt to comply with strict export restrictions
  • WCCFTech: NVIDIA Preps GeForce RTX 5090 DD For China As Export-Compliant Model, Reportedly Features Blackwell GB202-240 GPU

June 23: Reuters reported that DeepSeek is supporting China’s military and intelligence operations, based on an interview with a senior US State Department official. The official said DeepSeek was using “shell companies” in Southeast Asia to circumvent export restrictions. 

Reuters included comments from NVIDIA: 

“‘We do not support parties that have violated U.S. export controls or are on the U.S. entity lists,’ an NVIDIA spokesman said in a prepared statement, adding that ‘with the current export controls, we are effectively out of the China data center market, which is now served only by competitors such as Huawei.’” [...]

“‘Our review indicates that DeepSeek used lawfully acquired H800 products, not H100,’ an NVIDIA spokesman said, responding to a Reuters query about DeepSeek's alleged usage of H100 chips.” 

DeepSeek did not respond to an inquiry from Reuters. 

• Sources:
  • Reuters: Exclusive: DeepSeek aids China's military and evaded export controls, US official says
  • Asia Times: DeepSeek gets NVIDIA’s high-end GPUs via Singapore: US official

June 26: The Information reported that DeepSeek’s next AI model has been delayed due to a shortage of NVIDIA AI GPUs in China. 

• Sources:
  • The Information: DeepSeek’s Progress Stalled by U.S. Export Controls
  • Tom’s Hardware: AI disruptor DeepSeek's next-gen model delayed by NVIDIA GPU export restrictions to China — short supply of AI GPUs hinders development

July 2025

July 4: Bloomberg reported the US Department of Commerce is preparing a new export controls rule that would restrict the export of AI chips to Malaysia and Thailand. The rule’s goal would be to reduce AI chip smuggling to China. Based on its sources, Bloomberg said the export controls rule had not yet been finalized.  

• Sources: 
  • Bloomberg: US Plans AI Chip Curbs on Malaysia, Thailand Over China Concerns

July 10: Bloomberg reported that NVIDIA CEO Jensen Huang met with US President Donald Trump at the White House before traveling overseas to China. NVIDIA did not immediately respond to a request for comment from Reuters. 

• Sources: 
  • Bloomberg: NVIDIA’s Jensen Huang Meets with Trump Ahead of CEO’s China Trip
  • Reuters: NVIDIA CEO Huang to meet Trump before China trip, source says
  • CNBC: Trump hosts Jensen Huang at White House as NVIDIA tops $4 trillion market cap

July 11: In a public letter, a bipartisan group of US senators requested NVIDIA CEO Jensen Huang to avoid meeting with Chinese companies in an upcoming China trip that violate US laws or develop military applications that could undermine national security. 

Reuters included a response from NVIDIA about the senators’ letter: 

“An NVIDIA spokesperson said, ‘American wins’ when its technology sets ‘the global standard,’ and that China has one of the largest bodies of software developers in the world. AI software ‘should run best on the U.S. technology stack, encouraging nations worldwide to choose America,’ the spokesperson said.” 

• Sources: 
  • United States Senate letter to Jensen Huang
  • The Hill: Bipartisan senators press NVIDIA CEO over China trip
  • Reuters: US senators warn NVIDIA CEO about upcoming China trip
  • International Business Times: NVIDIA CEO Huang to Face Chinese Officials Over AI Export Curbs Just as Company Touches $4 Trillion Milestone

July 14: NVIDIA said it would soon resume sales of the H20 for customers in China. NVIDIA provided the following update in a blog post: 

“[Jensen] Huang also provided an update to customers, noting that NVIDIA is filing applications to sell the NVIDIA H20 GPU again. The U.S. government has assured NVIDIA that licenses will be granted, and NVIDIA hopes to start deliveries soon.”

The White House did not respond to a request for comment from CNN.

• Sources: 
  • NVIDIA Blog: NVIDIA CEO Jensen Huang Promotes AI in Washington, DC and China
  • CNN: NVIDIA says it will restart sales of a key AI chip to China, in a reversal of US restrictions
  • Reuters: Chinese firms rush to buy NVIDIA AI chips as sales set to resume
  • Associated Press: NVIDIA to resume sales of highly desired AI computer chips to China

The Malaysian government began requiring trade permits for all high-performance AI chips acquired from the United States. In a statement, the Malaysian government said: 

“The Ministry of Investment, Trade and Industry (MITI) would like to announce that, effective immediately, all exports, tranships and transits of high-performance AI chips of US origin are subject to a Strategic Trade Permit. These powers are provided for under Section 12 of the Strategic Trade Act 2010 (STA 2010), a Catch-All Control provision which requires individuals or companies to notify the relevant authority at least 30 days before exporting, transhipping, or bringing in transit any item not expressly listed in the Strategic Items List (SIL), if the individual or company knows or have reasonable grounds to suspect the item will be misused, or used for a restricted activity.

This initiative serves to close regulatory gaps while Malaysia undertakes further review on the inclusion of high-performance AI chips of US origin into the SIL of the STA 2010. Malaysia stands firm against any attempt to circumvent export controls or engage in illicit trade activities by any individual or company, who will face strict legal action if found violating the STA 2010 or related laws.”

• Sources: 
  • Malaysia Government Release
  • Reuters: Malaysia says trade permit required for AI chips of U.S. origin
  • Bloomberg: Malaysia Controls AI Chip Exports As US Targets China Smuggling

July 15: DigiTimes reported that NVIDIA is preparing a new AI GPU for the China market, the RTX 6000D. DigiTimes claimed the card would become available in the third quarter of 2025, according to its sources in the supply chain. 

• Sources: 
  • DigiTimes: Exclusive: Jensen Huang's third visit to China in 2025; RTX 6000D aims for two million shipments
  • Tom’s Hardware: NVIDIA reportedly preparing RTX 6000D for Chinese market to comply with U.S. export controls — fabricated on TSMC N4, featuring GDDR7 memory capable of delivering 1,100 GB/s of bidirectional bandwidth

July 24: The Financial Times reported that more than $1B worth of NVIDIA’s AI chips had been smuggled to China. In response, NVIDIA said that building datacenters with “smuggled products” was a “losing proposition.” 

• NVIDIA statement: 

• “Trying to cobble together datacenters from smuggled products is a losing proposition, both technically and economically. Datacenters require service and support, which we provide only to authorized NVIDIA products.” - NVIDIA to CNBC

• Sources:
  • Financial Times: NVIDIA AI chips worth $1bn smuggled to China after Trump export controls
  • CNBC: NVIDIA addresses AI chip smuggling, says bootleg data centers are a ‘losing proposition’

July 28: The Financial Times reported that the US Commerce Department was not going to make “tough moves” to tighten export controls to China. According to the report, the US government would try to secure a better trade deal with China ahead of negotiations in Stockholm.

The Washington Post reported that several congressional members had warned the US administration against loosening its export controls for AI GPUs. NVIDIA and the US Commerce Department did reply to requests for comment to The Washington Post.

Several national security experts voiced their concern by sending a letter to the US Commerce Department. 

• Sources: 
  • Financial Times: Donald Trump freezes export controls to secure trade deal with China
  • The Washington Post: Trump’s retreat on China chip ban triggers policy spat
  • Tom’s Hardware: Trump freeze on export restrictions to China reportedly in aid of trade talks — White House seeking face-to-face with Xi Jinping as dissenters warn H20 reversal is a dangerous mis-step

July 29: Reuters reported that NVIDIA had ordered 300,000 more H20 chips from TSMC due to strong demand from its customers in China. Several weeks prior, NVIDIA said it would resume sales of the H20 chip to China.

NVIDIA declined to comment to Reuters.

• Sources: 
  • Reuters: Exclusive: NVIDIA orders 300,000 H20 chips from TSMC due to robust China demand, sources say
    • Repost in MSN
• Hot Hardware: TSMC Secures 300,000 H20 AI Chip Order As NVIDIA Boosts Supply To China

July 31: The New York Times reported that Chinese government officials asked NVIDIA for information about security risks associated with its H20 chip. NVIDIA denied having “backdoors” in its chips. 

• NVIDIA statement: 

• “Cybersecurity is critically important to us. NVIDIA does not have ‘backdoors’ in our chips that would give anyone a remote way to access or control them.” - NVIDIA to CNBC

• Sources: 

• The New York Times: China Summons NVIDIA Over ‘Backdoor Security’ Risks of A.I. Chips
• CNBC: NVIDIA denies its China-bound H20 AI chips have ‘backdoors’ after Beijing’s security concerns
• Reuters: NVIDIA says its chips have no 'backdoors' after China flags H20 security concerns
• Associated Press: China summons NVIDIA over ‘backdoor safety risks’ in H20 chips

August 2025

August 4: A government official told Bloomberg the United States is exploring adding location trackers for AI chips. Bloomberg quoted the official as saying, “There is discussion about potentially the types of software or physical changes you could make to the chips themselves to do better location-tracking.”

• Sources: 
  • Bloomberg: US Explores Location Trackers for AI Chips, Official Says
  • The Register: Uncle Sam floats tracking tech to keep AI chips out of China

August 5: In a blog post, NVIDIA said that its GPU products do not have backdoors or kill switches. 

• Sources: 
  • NVIDIA Blog: No Backdoors. No Kill Switches. No Spyware.
  • CNBC: NVIDIA says its AI chips don’t have a ‘kill switch’ after Chinese accusation

The US Department of Justice announced it had arrested two people in California for smuggling high-end GPUs to China that purportedly amount to “tens of millions of dollars’ worth of sensitive microchips used in artificial intelligence (AI) applications.” The BBC reported that court documents say the shipments included the NVIDIA H100 and RTX 4090. 

• NVIDIA statement:

• “This case demonstrates that smuggling is a nonstarter. We primarily sell our products to well-known partners, including OEMs, who help us ensure that all sales comply with U.S. export control rules. Even relatively small exporters and shipments are subject to thorough review and scrutiny, and any diverted products would have no service, support, or updates.” - NVIDIA to TechCrunch

• Sources: 
  • Department of Justice release
  • Bloomberg: US Charges Chinese Nationals With NVIDIA Chips Export Breach
  • Reuters: Two Chinese nationals in California accused of illegally shipping NVIDIA AI chips to China
  • New York Post: Chinese nationals living in US charged with smuggling millions worth of NVIDIA’s powerful AI chips to Beijing
  • TechCrunch: Two arrested for smuggling AI chips to China; NVIDIA says no to kill switches
  • BBC: Chinese nationals charged with exporting NVIDIA AI chips to China

August 10: The Financial Times reported that NVIDIA would give the US government 15% of its revenue from H20 chip sales from customers in China. The deal is reportedly part of an agreement that would allow NVIDIA to acquire export licenses from the Commerce Department in order to sell the H20 chip in China. AMD would be subject to the same rules for the MI308. 

• NVIDIA statement: 

• “We follow rules the U.S. government sets for our participation in worldwide markets. While we haven’t shipped H20 to China for months, we hope export control rules will let America compete in China and worldwide. America cannot repeat 5G and lose telecommunication leadership. America’s AI tech stack can be the world’s standard if we race.” - NVIDIA to the Associated Press

• Sources:
  • Financial Times: NVIDIA and AMD to pay 15% of China chip sale revenues to US government
  • The Associated Press: NVIDIA and AMD to pay 15% of China chip sale revenue to US government in an unusual agreement
  • The New York Times: U.S. Government to Take Cut of NVIDIA and AMD A.I. Chip Sales to China
  • BBC: NVIDIA and AMD to pay 15% of China chip sales to US

August 11: According to a report in Bloomberg, US President Trump said he was open to allowing NVIDIA to sell a modified Blackwell chip for the China market. The US President also said that he has negotiated with NVIDIA CEO Jensen Huang about the deal to allow H20 sales in China. 

• Sources:
  • Bloomberg: Trump Open to NVIDIA Selling Scaled-Back Blackwell Chip to China
  • The Register: Trump seeing green as he weighs deal to allow NVIDIA Blackwell GPU sales to China
  • NPR: Trump says NVIDIA will hand the U.S. 15% of its H20 chip sales to China
  • Reuters: Trump opens door to sales of version of NVIDIA’s next-gen AI chips in China
    • Repost in MSN

August 12: Bloomberg reported that Chinese officials had “urged local companies” to avoid purchasing and using NVIDIA’s H20 chip, especially for national security and government work. According to Bloomberg, China questioned companies whether they had found security problems with NVIDIA’s chips.  

Bloomberg included commentary from NVIDIA: 

“AMD declined to comment, while NVIDIA said in a statement that ‘the H20 is not a military product or for government infrastructure.’ China has ample supplies of domestic chips, NVIDIA said, and ‘won’t and never has relied on American chips for government operations.’

• Sources:
  • Bloomberg: China Urges Firms Not to Use NVIDIA H20 Chips In New Guidance
  • Reuters: China cautions tech firms over NVIDIA H20 AI chip purchases, sources say

CNBC reported that the Trump Administration was still working on the details for how to implement the 15% export tax on NVIDIA and AMD for selling certain chips to China. 

• NVIDIA statement:

• “We follow rules the U.S. government sets for our participation in worldwide markets.” - NVIDIA to CNBC

• Sources:
  • CNBC: White House says it’s working out legality of NVIDIA and AMD China chip deals
  • Tom’s Hardware: White House confirms it's still figuring out the legality of the revenue-sharing NVIDIA and AMD deal for China GPU sales — 'The legality of it, the mechanics of it, is still being ironed out'

August 13: Reuters reported that US officials have covertly placed “location-tracking devices” in targeted shipments with advanced chips in an effort to catch chip smuggling to China. Unnamed sources told Reuters that the tracking devices had been placed in shipments of OEM servers, including from Dell and Supermicro. 

NVIDIA declined to comment to Reuters. 

• Sources: 
  • Reuters: Exclusive: US embeds trackers in AI chip shipments to catch diversions to China, sources say
  • Tom’s Hardware: U.S. authorities allegedly placed secret tracking devices in AI chip shipments to China — report claims targeted shipments from Dell and Super Micro containing NVIDIA and AMD chips had trackers in packaging and servers themselves

TweakTown: US authorities secretly place location tracking devices in targeted AI chip shipments to China

Our review of the AMD RX 9060 XT 16GB tests the card against the NVIDIA RTX 5060, RTX 5060 Ti 16GB, RX 7700 XT, 6600 XT, and similar GPUs

The Highlights

• The RX 9060 XT comes with either 8GB or 16GB of VRAM
• The 8GB variant has VRAM limitation issues
• AMD, like NVIDIA, is taking advantage of consumers more aggressively than we've seen in the past
• Original MSRP: $350 (16GB) and $300 (8GB)
• Release Date: June 5, 2025

Table of Contents

• AutoTOC

Intro

Quick version up front: Against the RTX 5060 (read our review), the RX 9060 XT 16GB is anywhere from 2% to 17% better at 1080p rasterized, typically 10%, the same to 22% better at 1440p, and worse to 21% better at 4K, with more variability here. In RT, the 9060 XT again ranges from significantly worse than the RTX 5060 in Black Myth: Wukong to can’t-divide-by-zero better in some specific tests where the 5060 ran out of VRAM. This would also happen to the 9060 XT 8GB model, which…exists. Against itself, AMD’s ray tracing performance has continued to show significant improvement over the last generation, with it sometimes equating the 7700 XT (read our review) in RT scenarios.
Now that we have the important stuff out of the way, before we get started, a quick callback to a few months ago.

Editor's note: This was originally published on June 4, 2025 as a video. This content has been adapted to written format for this article and is unchanged from the original publication.

Credits

Test Lead, Host, Writing, Editing

Steve Burke

Testing

Patrick Lathan

Testing, Editing

Mike Gaglione

Camera, Video Editing

Tim Phetdara

Camera

Andrew Coleman

Camera, Testing

Tannen Williams

Writing, Web Editing

Jimmy Thang

Unfortunately, we’re back to that because AMD can’t just launch a product and shut the f*ck up. It has to shove its foot so deep into its own mouth that it comes out the other side. 

Today, we’re reviewing AMD’s RX 9060 XT. The 9060 XT comes in 8GB and 16GB options, but beyond this, it comes with the usual side of AMD executive tweets.

AMD’s Frank Azor and guy who once bet Twitter $10 that he’d have better supply than NVIDIA during COVID said this:

“Majority of gamers are still playing at 1080p and have no use for more than 8GB of memory. Most played games WW are mostly esports games. We wouldn’t build it if there wasn’t a market for it. If 8GB isn’t right for you then there’s 16GB. Same GPU, no compromise, just memory options.”

First of all, AMD is the king of building things that there’s no market for. For example, these include:

• The 2004 Personal Internet Communicator it abandoned in under 2 years
• Bulldozer
• Anti-Lag+ the first time, before pulling it because it got people VAC banned
• The FX-9370
• The Radeon Pro Duo with almost no actual market
• Athlon 64 X2 3800+ EE SFF
• Radeon VII
• Vega Frontier Edition
• Vega 64
• AMD RX 6300
• The Radeon 6500 XT with PCIe Gen 4 x 4
• The HD 2900 XT
• AMD Fusion Render Cloud in 2009
• AMD Live!
• And finally, the AMD mountain bike, which is still the proudest we’ve ever been of a product that GN killed entirely on its own.

Secondly, Azor wrote, “same GPU, no compromise” while having half the VRAM is literally a compromise. That is the definition of a compromise in the same way Corsair’s i500 was such a compromise as an SFF PC that it removed all the “no compromise” language after our review. Rebranding “compromise” as “options” is the type of thing you put on your job application to become a Chief Architect of Marketing. 

Oh… OK, bad example.

In June 2020, AMD posted a blog post saying that 4GB of VRAM “is evidently not enough for today’s games.” In January 2022, AMD hides the blog post just ahead of the launch of the RX 5500 XT 4GB GPU, then panicked when it was discovered and unhid the blog post.

In 2020 after NVIDIA’s paid 8K marketing campaign, Azor tweeted “I guess we’ll scratch our 16K gaming message” in a jab at NVIDIA. 

In November 2020, AMD partners with Sony to slap a giant “8K” sticker on the PlayStation 5 boxes that it later removed the label from.

Also in November, 2020, Frank Azor tweeted about successfully buying his own product in a market with a massive GPU shortage presumably to save $10 from a bet.

In 2017, AMD posted a video that said “Poor Volta” before having its ass handed to it by NVIDIA for the next several generations.

Also in years past, the Fury is called the “overclocker’s dream.”

In 2016, AMD former Corporate Vice President Roy Taylor sat next to us at the RX 480 (watch our review) launch event and looked at us when we were looking up the Fury X (read our review) specs. He asked why we needed them. We said because we were trying to remember if it was 4GB of memory, and he said, “we’ll never make that mistake again.”

That’s the same launch event that had two RX 480s in CrossFire to show them beating a single GTX 1080 in Ashes of the Singularity, even though AMD’s marketing team ran the CPU benchmark for one brand and the GPU benchmark for the other in a completely incomparable benchmark.

AMD is single-handedly keeping Podithodontists employed by requiring so many emergency foot-from-mouth extraction surgeries.

The point is that an 8GB card should not exist for $300 today. Not for NVIDIA, not for AMD. 8GB cards do have a place -- he’s right about that. But it’s not in $300 solutions. We had 8GB cards for $200 years ago when the RX 580s (read our review) were on fire sale. It’s as if VRAM has stopped advancing, and it is actively hurting gaming today. You’ll definitely notice this within a few years, but it’s also likely that it’ll affect the gaming experience for many people today.

If AMD wants to sell an 8GB “esports card” or whatever the company wants to call it, it can do that -- but do it for the true low-end. AMD is just right back to playing NVIDIA minus $50. AMD has to invest in its brand, like Intel is doing with Arc, where the company is willing to take less money to establish market share to try and eventually grow longterm. AMD could also shut the f*ck up on Twitter, specifically getting the executives off the social media platform. It’s not helping. The investments are supposed to go into products and getting market share, not into one-upping NVIDIA’s f*ckups. Giving out review samples doesn’t undo weaponization of incompetence from AMD marketing.

GPU Test Bench

RX 9060 XT Thermals

We’ll run a thermal test to evaluate how the Sapphire Pulse model does. These lower-end cards sometimes get shafted with poorly applied thermal solutions or just mounting issues.

In a 100% rendering workload, the 9060 XT GPU temperature sensor climbed to about 56 degrees Celsius steady state. That’s completely acceptable. The hot spot temperature was 79 degrees, meaning that there’s about a 23-degree delta between the two. The highest we’ve seen for GPUs has been in the 40s for a poorly installed GPU cooler, which is why we pay attention to the delta to begin with. In this instance, 23 is higher than we’d like to see, but both numbers are well within tolerances. We can thank the relatively low power limit of this particular GPU for that more than the cooler.

Memory temperatures ran warm at 85 degrees Celsius in an ambient of 21C. We’d like to see this cooler as it leaves little room for aging and ramping internal case ambient temperatures, but there’s still enough room here that it’s OK overall. We evaluate this closer in our tear-down video.

One thing we noticed was the bump right at the start: The fans took a little bit to kick in and didn’t hit until the GPU bounced off of what we assume is the VBIOS temperature target of 56 degrees for the GPU. After this, they aggressively cooled down the card, then slowly ramped back to that temperature target.

Frequency Target

Next, we checked the frequency log to ensure the card is hitting the advertised clock speed. In the same test where the GPU leveled-out at 56 degrees Celsius, the core frequency held at about 3130 MHz or so. The advertised boost frequency is “up to” 3130 MHz according to AMD, so it’s hitting that marker.

RX 9060 XT Acoustics

We also ran a quick acoustic test in our hemi-anechoic chamber that we built. The RX 9060 XT Pulse ended up almost being too quiet to get a good read on, as it’s near our noise floor. The device measured at about 16 dBA under full load and tested at 1 meter distance in the chamber, shown by the green line in this plot with some 9070 (read our review) models for reference.

The 9060 XT Pulse ended up with a slight spike at 500 Hz, which we’ve seen in the other cards on this chart, and then otherwise a relatively gradual frequency falloff later on in the plot. Nothing jumped out at us here, which is a compliment for the Pulse. The low power draw of the card allows it to spin relatively slowly. It is likely that the other fans in your system would be louder than this GPU.

RX 9060 XT Game Benchmarks

Dragon’s Dogma 2 - 4K

Dragon’s Dogma 2 is up first. Tested at 4K, the RX 9060 XT outperforms the RTX 5060 GPU. We’re in unplayable territory, but this is still a useful tool to understand scaling across resolutions.

The RX 9060 XT leads the 5060 by 7% here, with both behind the former flagship 2080 Ti GPU. The jump from the 9060 XT to the 9070 is huge, at 79% improved to the next card up. AMD has a large performance gap between these models. In between, we can find the 7900 XT (read our revisit) and 7800 XT (read our review) also from AMD, with the 6950 XT around the levels of the 7900 GRE (read our review). NVIDIA competition additionally includes the 5060 Ti (read our review), which ran at 40 FPS AVG and outperformed the 9060 XT by 11%.

Dragon’s Dogma 2 - 1440p

At 1440p, the RX 9060 XT ran at a playable 65 FPS AVG with lows proportionately spaced. The lows for all the latest NVIDIA and AMD cards are indicative of consistent frametime pacing for this game.

The 65 FPS result positions the 9060 XT near the level of the 7700 XT and 3070 Ti, just below the 5060 Ti. The 5060 Ti has a 7% lead, with the 9060 XT leading the 5060 non-Ti by 14%. The 9070 is significantly better, at 106 FPS AVG to the 9060 XT’s 65 FPS AVG, so there is a large gap between these. Currently, that gap is filled by the 7800 XT, 7900 GRE, and 6950 XT of prior generations from AMD, or the 3080, 4070, and 5070 (read our review) from NVIDIA.

Dragon’s Dogma 2 - 1080p

At 1080p, the 9060 XT again sits around the 3070 Ti and 7700 XT levels of performance. That positions it ahead of the RTX 5060 by 13%, with the 5060 Ti ahead of the 9060 XT by 8%. The 9070 improves to 134 FPS AVG from the 86 FPS AVG of the 9060 XT, or 56%, with the 9070 XT (read our review) improving by 68.5% to 145 FPS AVG, just below the 5070 Ti and prior XTX flagship.

Generationally and by name only (which, again, has shifted in meaning for both AMD and NVIDIA), the 9060 XT improves on the 7600 (watch our review) by 42% and the 6600 XT by 48%. The 6600 XT launched at around $380, but was available later in its life commonly for $220 to $260.

FFXIV 4K

Final Fantasy 14: Dawntrail is up now, first at 4K.

By framerate alone, the RX 9060 XT is basically the exact same product as the RX 6700 XT here. The two are nearly identical in performance, though there is a spec difference, most notably with VRAM. The original MSRP for the 6700 XT was around $480, and later in its life it was often available for around $300 to $350. In other words, against only the launch price, AMD has come down about $130 to equal the performance of a card from 4 years ago. Listings for a used 6700 XT range from about $250 (sold) to about $400, with common options in the $280-$320 range.

The 4060 Ti is also about tied with the 9060 XT, with the 7700 XT slightly ahead. The 5060 leads the 9060 XT in this game and benchmark, part of our established fact that AMD just generally struggles in this benchmark right now as compared to NVIDIA. In fact, even Intel’s B580 (read our review) is outperforming the 9060 XT. In this test, the 9060 XT just doesn’t do that well overall. The same was true of the 9070 and 9070 XT in this testing.

FFXIV 1440p

At 1440p, the 9060 XT ran at 85 FPS AVG, which has it equivalent to the 6700 XT from 2021, slightly below the B580 (although with better 0.1% lows), and just barely ahead of the RTX 4060 Ti. This game is one of Intel’s more promising for performance gains against the incumbents and remains one of AMD’s weaker games.

The 5060 leads the 9060 XT by 2% for AVG FPS, with lows functionally the same. The RTX 5060 Ti at 104 FPS AVG leads the 9060 XT by 22%. As we’ve seen in the past, AMD’s competitive performance in FFXIV just isn’t as strong as in other titles.

Generationally, the 9060 XT leads the RX 6600 XT’s 63 FPS AVG by 36%, the 6600 by 62%, and NVIDIA’s once most common GPU, the GTX 1060 6GB, by 158%.

Starfield - 4K

Starfield is up next. Sometimes, people ask why we still test this game. Well, we figured it made sense to benchmark a game nobody plays with cards nobody wants to buy.

AMD’s RX 9060 XT ran at 39 FPS AVG, with lows proportional to the average. None of the NVIDIA or AMD cards on this chart have particularly strong or weak 1% and 0.1% lows. The RTX 5060 Ti 16GB at 41 FPS AVG outperforms the 9060 XT 16GB by 5%. 

Against the new RTX 5060 that we bought for $330, the RX 9060 XT leads by 18.8% in average framerate. Lows scale along with the average. We need to drop resolution for more cards to compare.

Starfield - 1440p

Dropping to 1440p, the RX 9060 XT ran at 62 FPS AVG and ended up between the 5060 Ti and 3070 Ti. The RTX 5060 Ti’s 65 FPS AVG is indistinguishably different from the 9060 XT from an experiential standpoint, but technically ahead by 6%.

Buying a higher-end card instead, like the 9070, would improve the performance in this title by 56%, from 62 FPS AVG to 96 FPS AVG. One of the cheapest RX 9070s in stock at the time of writing was $650, about $100 over MSRP, which means that a $350 9060 XT 16GB (if you could find it at that price) would achieve 64% of the performance for 54% of the price. We expect it’ll be more than $350.

Down the stack and against older generations, the 9060 XT distances itself from the 6700 XT this time (despite similar performance in other games), and improves upon the 6600 XT by 63%, the 6600 by 94%, and the 2060 by about the same. Intel’s B580 is similar to the RX 7600 in performance, but suffers in frametime consistency in this game. Its experience would be noticeably worse as a result.

Starfield - 1080p

Tested at 1080p, the 9060 XT ran at 77 FPS AVG and maintained consistent lows, just like the other AMD and NVIDIA cards at its flanks. The 5060 Ti runs just 4.7 FPS AVG higher framerate, or a 6% advantage. Against the 5060 non-Ti, the 9060 XT leads by 17%. 

AMD’s prior 7900 GRE, 6950 XT, and NVIDIA’s new 5070 all end up ahead of the 9060 XT in a similar clustered ranking. Beyond that, the 9070’s 118 FPS AVG is 52% ahead. By MSRP, which the 9070 still regularly isn’t, that’d be a 52% performance gain for a 57% price increase by MSRP, but we can’t be sure where the 9060 XT will land until launch.

Resident Evil 4 - 4K

Resident Evil 4 is up now. At 4K, the 9060 XT’s 56 FPS AVG has it about equal to the 5060 Ti. Neither has a frametime consistency advantage over the other and all of these metrics are within variance. They’re the same result. The 3070 Ti is just behind the 9060 XT, with the 7700 XT just ahead of the 5060 Ti and 9060 XT. AMD improves upon the RTX 5060’s 46 FPS AVG by 20%, which is the same improvement from the Intel A770 (read our revisit). In this game, the Arc GPUs do better for frametime consistency than they have elsewhere. The 9060 XT is also 22% improved over the Intel B580, although the B580 is commonly cheaper.

Generationally, noteworthy markers in this chart include the RTX 2060, RX 6600, RX 6600 XT, and RX 7600, all of which would be significantly improved upon with a 9060 XT or 5060 Ti-class GPU.

Resident Evil 4 - 1440p

At 1440p, the RX 9060 XT held a 106 FPS AVG, putting it between the 5060 Ti and 3070 Ti. The 5060 Ti is 6% ahead in average framerate. The 7700 XT and 4070 outperform both of these, with the 5070 notably ahead both in price and performance.

The 9070 again leaves a large gap between it and the 9060 XT, with a 62% improvement over the new card. Versus the 5060, the 9060 XT runs 17% ahead, about the same as its uplift over the 3070, 4060 Ti, and similar to the 6700 XT. Intel’s B580 does OK to hang on here, but isn’t keeping up with the newer 5060 and 9060 XT. It’s getting closer though, so we’ll see what Intel achieves in the next generation.

Versus older cards, users of the 2060 would see an improvement of 127% to the 9060 XT, 6600 owners would see the jump from 58 FPS to 106 FPS AVG, nearly doubling, and 6600 XT users would also see meaningful uplift.

Black Myth: Wukong - 4K

Black Myth: Wukong is up now, tested first at 4K just for scaling purposes. 

The 9060 XT ran at 25.6 FPS AVG here. That’s obviously not playable, but let’s look at scaling: Against the 5060, performance is about the same. The 5060 Ti is 22% ahead, so similar to what we’ve seen even in higher framerate scenarios. The 9070 improves on the 9060 XT by 61% here. Let’s move to a lower resolution.

Black Myth: Wukong - 1440p

At 1440p, the 9060 XT climbs to 49 FPS AVG. That has it at about the same level as the 7700 XT and 3070 Ti. The 5060 Ti is about 17% ahead of the 9060 XT in this test, with the 7800 XT similarly ahead. As for the new 5060, that provides functionally the same experience when not running into VRAM issues. The 3070 is around the same spot as the 5060.

The 9060 XT at least improves over prior generation parts in a meaningful way: Against the RX 6600 XT’s 29 FPS AVG, we saw a 66% improvement. Versus NVIDIA’s old and popular RTX 2060, the uplift is 94% from 25 FPS AVG.

Black Myth: Wukong - 1080p

1080p puts the 9060 XT into playable territory without any form of upscaling and with these quality settings. The 9060 XT ran at 71 FPS AVG with lows where we’d expect them. The 5060 Ti’s 81 FPS AVG has it 14% ahead of the 9060 XT here, with the 5060’s 66 FPS AVG giving up an 8% lead to the 9060 XT. This is a stronger title for NVIDIA even when rasterized.

Intel’s B580 does OK for frametime pacing this time, but ends up far down the stack and adjacent to the old RTX 3060 and prior generation RX 7600. It’s just not competitive here.

Dying Light 2 - 4K

Dying Light 2 is up now. 

4K isn’t worth much time: First of all, as we all know, the human eye can only see 30 FPS without running two of them in SLI, which enables 60 FPS visibility. One of our viewers with glasses in the last review pointed-out that his spectacles enable MFG 4X, multiplying the visible framerate by a further 2X on top of SLI eyeballs.

All this to say that, at 32 FPS AVG for the 9060 XT, clearly we’re past perfection anyway.

The 5060 Ti outperforms the 9060 XT by 20% here, similar to some of the other tests even in spite of the heavy workload. That’s why we include these tests, though. The 5060 is about the same. Let’s move on.

Dying Light 2 - 1440p

At 1440p, the 9060 XT ran at 63 FPS AVG with good lows -- but not any better than any of its competitors’. The 3070 is slightly ahead and, critically, the B580 is actually about the same. Intel does well in this one, outperforming the 5060 in average and lows, while being functionally equal to the 9060 XT’s metrics. That’s more exciting than NVIDIA or AMD here.

But as for NVIDIA: The 5060 and 9060 XT would feel about the same in play in this game, with the 5060 Ti improving on the 9060 XT by 17%. Users of the now venerable RTX 3080 10GB can feel pretty good about their purchase, because the card is still hanging in there and keeping pace with modern GPUs, not too distant from the RX 9070.

Speaking of, the 9070’s uplift over the 9060 XT is 68%, putting them in totally different price and performance classes.

Cyberpunk 2077: Phantom Liberty - 4K

Cyberpunk 2077: Phantom Liberty is up next.

Tested at 4K, the RX 9060 XT ran at about 30 FPS AVG, so although there’s only a 6% lead for the 5060 Ti, we’re too constrained in performance to have a full picture yet. 

The 5060 fell notably below the 9060 XT here, with the AMD device holding a 21% advantage. Let’s see if that sticks at 1440p.

Cyberpunk 2077: Phantom Liberty - 1440p

At 1440p, the 9060 XT ran at 64 FPS AVG and landed between the 3070 Ti and 5060 Ti. The lead over the 5060 non-Ti drops to 16.7% here, down from about 21% at the less playable 4K. As for the 5060 Ti, its lead actually didn’t change much: It’s at 7.5% ahead now, from around 6% previously.

The 9060 XT is playable with these settings. Moving to something like a 5070 or 9070 would obviously be a huge jump up and would give some more room for higher settings or just future games, or even mods in this game, but that’s also a big price jump.

Cyberpunk 2077: Phantom Liberty - 1080p

At 1080p, the 9060 XT pushed up to a 100 FPS AVG. The 5060 Ti’s lead is now 8.8%, so we’ve seen it slowly increase its advantage over the 9060 XT as resolution has decreased. That’s why we run the 4K tests even when performance is too limited to play. As for the 5060, the 9060 XT leads it by 12% now, so it has consistently dropped from the 4K lead of about 21%. 

RX 9060 XT Ray Tracing Benchmarks

Ray tracing benchmarks are next.

Ray Tracing - Black Myth: Wukong 4K

We’ll start with the worst case scenario for everyone except NVIDIA, which is Black Myth: Wukong.

First at 4K upscaled, the 9060 XT ends up at the bottom of the stack. AMD has at least improved its ray tracing performance over the 7000 series, shown with the 7900 XT equaling the 9060 XT, but it’s not good enough to contend with NVIDIA (even if ignoring the low framerate at this resolution). The 5060 runs a 52% higher average framerate than the 9060 XT here. Let’s move to something that might actually run.

Ray Tracing - Black Myth: Wukong 1440p

At 1440p upscaled and ray traced, the 9060 XT ran at 31 FPS AVG and landed just ahead of the 2080 Super. That gives the 5060 non-Ti and its 45 FPS AVG a lead of 44%, down from the 52% previously. We’ll see if that continues to the more playable next resolution. The 5060 Ti ran at 54 FPS AVG, leading the 9060 XT by 72%. Even the 9070 XT was technically below the 5060 Ti, showing just how this particular game remains undefeated for NVIDIA.

Ray Tracing - Black Myth: Wukong 1080p

At 1080p upscaled, the 9060 XT ran at 47 FPS AVG. As always, for us, it’s not about the absolute framerate but instead about the relative framerate. In a relative sense, the 5060 ends up 37% ahead of the 9060 XT. In an absolute sense, it’s also just at a more playable FPS. The interesting part is the shrinking of the gap at lower resolutions here: Possibly because of how overrun it was, AMD’s GPU went from yielding a 52% advantage to the 5060, to 44% at 1440p, to 37% at 1080p. At this rate, maybe they’ll be at parity at 144p or something.

The 5060 Ti ends up 60% ahead here, with the 9070 XT around the same level. AMD gets left behind in this particular game with ray tracing on.

Ray Tracing - Dragon’s Dogma 2 4K

Dragon’s Dogma 2 with RT is up next. At 4K and not upscaled, just at normal 4K, the 9060 XT ran at 32 FPS AVG, basically tying the 2080 Ti and just behind the 3070 Ti. The 5060 Ti leads by 11% here, with the 9060 XT leading the 5060 non-Ti by 6.4%. Better performers include the 5070 at 49 FPS AVG, or 53% improved, and the 9070, at 76% improved. That breaks rank from what we saw in most rasterization tests. Of course, you could keep scaling up and we have results for that as well.

Ray Tracing - Dragon’s Dogma 2 1440p

At 1440p with RT and no upscaling, the 9060 XT’s 56 FPS AVG puts it well into playable territory with these settings. That has the 7700 XT just barely ahead, but the same from an experiential standpoint, and the 3070 is behind. The lead over the 5060’s 50 FPS AVG is 12% for a somewhat meaningful advantage. The 5060 Ti has a similar lead of 10% over the 9060 XT.

Compared to prior generations, AMD has improved on its architecture significantly and boosted its RT performance meaningfully. This is something we already saw in the 9070 XT review, but it continues at the lower end: The 9060 XT is a big jump over cards like the 6600 XT and 6600 previously.

Ray Tracing - Dragon’s Dogma 2 1080p

At 1080p, the 9060 XT’s 78 FPS AVG put it between the 7700 XT and 5060 Ti. The 5060 Ti’s lead drops from the 10% we saw at 1440p to just 5.5% at 1080p, with the 9060 XT increasing the gap against the 5060, now at 17% from 12% before. AMD is scaling better in this test as resolution decreases, even with ray tracing enabled.

Ray Tracing - Dying Light 2 1440p

We’re moving on to Dying Light 2 with ray tracing at 1440p upscaled. 

The 9060 XT was at 51 FPS AVG here and didn’t have any issues to speak of for frametime consistency. That has it ahead of the 5060 by less than 2 FPS, which wouldn’t be noticeable in real play. The 5060 Ti’s lead is more meaningful, at 17.8% over the 9060 XT. Overall, AMD is mostly improving on its own past performances here: Nearly hitting 7800 XT levels with the 9060 XT so at least that’s an improvement in RT.

Ray Tracing - Dying Light 2 1080p

At 1080p, the 9060 XT’s 77 FPS AVG has it again at about the level of the 7800 XT, just ahead of the 7700 XT. The 5060 Ti is about 13% better in average framerate, with the 9070 48% better than the 9060 XT and similar to the 5070.

AMD’s 9060 XT is 7.8% ahead of the RTX 5060 non-Ti here. The B580 is down below that, though not far below, and has OK frametime pacing this time. The 3060 Ti remains a weirdly relevant comparison to the 5060, which sort of feels like a 3060 Ti Super or something.

Ray Tracing - Resident Evil 4 4K

Resident Evil 4 ray traced is next, first at 4K upscaled. The RX 9060 XT 16GB ran at 63 FPS AVG here, putting it relatively close to the 5060 Ti’s 67 FPS AVG. The 5060 Ti has a lead of just 7% here, which is good for the 9060 XT. The 5060 gives the 9060 XT a 14% advantage, so it’s positioned better in this test than some of the other RT benchmarks. Lows are also consistent and proportional to the average, similar to the flanking devices from NVIDIA and AMD. Intel actually also had an OK overall performance here, with a 50 FPS AVG that begins to threaten the RTX 5060, though not the 9060 XT. Intel is mostly contending with the fact that it’s not consistently competitive, despite having improved this significantly with Battlemage.

Ray Tracing - Resident Evil 4 1440p

At 1440p upscaled and with RT, the RX 9060 XT’s 102 FPS AVG puts it just below the 7700 XT. The 5060 Ti is now 10% ahead, up from its 7% lead at 4K. The 9060 XT’s 10 FPS lead over the RTX 5060 non-Ti also reduces its advantage to 11% from 14% previously.

Ray Tracing - C77 VRAM Issues (1080p, RT Medium)

Finally, we’ll show the VRAM issues on the 8GB RTX 5060. These will also apply to the RX 9060 XT 8GB model.

In this ray tracing benchmark of Cyberpunk 2077: Phantom Liberty with 1080p/RT Medium settings, the RTX 5060 ran at 51 FPS AVG -- which sounds OK -- but had awful lows at 21 FPS for 0.1%. What’s important is looking at its neighbors: The RX 7800 XT had nearly the same average, but its lows indicated far more consistent and less choppy frametime pacing with the 43 FPS 0.1% result. The 9060 XT held a 54 FPS AVG and had well-timed, consistent frames, resulting in good 1% and 0.1% lows that represent an overall consistent experience. 8GB cards struggle here today, already, as shown with the 5060 and RTX 3070 Ti.

Ray Tracing - C77 VRAM Issues (4K, RT Medium)

To illustrate the point, here’s 4K with RT Medium. It’s ridiculous, but that’s the point: Despite struggling hard and overall being stuttery with its 15 FPS AVG, the 9060 XT had, to its credit, excellently paced frametimes for its average at 13.7 FPS and 13.6 FPS. There’s not a ton of data to work with here at this framerate, but the point is that even this sh*t experience is infinitely more playable than what we observe on cards like the RTX 3070 Ti, 3070, and RTX 5060, all of which are so variable from exceeding VRAM that the numbers are meaningless. We really can’t tell them apart: They’re all just unplayable in more ways than the 9060 XT is here.

So this is an illustration of VRAM limitations in real-time.

RX 9060 XT Efficiency Benchmarks

Up next, we’ll look at efficiency benchmarks when capturing GPU power consumption via the PCIe cables and the PCIe slot combined. This is isolated power draw to just the GPU. We’re representing efficiency in the form of FPS/W, or frames per joule, with higher being more efficient. The product name also contains the power drawn during the specific test.

Efficiency: Starfield 1440p

In Starfield at 1440p, the RX 9060 XT ended up slightly more efficient than the 9070 XT, though obviously with a lower framerate. The 9060 XT pulled 169W in this test, resulting in a 0.36 FPS/W result. The 4060 has a much lower framerate that’d push it into lower graphics settings, but is more efficient and has lower overall power draw. The RTX 5060 was also behind the 9060 XT in framerate, but also ahead in efficiency as a result of reduced total power draw of 126W for its framerate. The 5060 Ti pulled 145W and had a higher framerate than the 9060 XT, allowing it to approach the top of the charts with the 4060 Ti. Both of these are the 16GB models.

Efficiency: F1 24 RT 4K

F1 24 with RT is next and at 4K. The RTX 5060 couldn’t run this test without major stuttering, so it’s not on the charts.

The 9060 XT had a 0.17 FPS/W result, with the 5070 tying it also at 0.17 FPS/W, but with a 246W power draw. The 5060 Ti pulled 176W here. The 9060 XT is about the same power consumption as the 5060 Ti and only slightly behind in framerate, as this is one of AMD’s stronger RT titles. As a result, the efficiency is almost the same.

Efficiency: F1 24 RT 1080p

At 1080p with RT, the 9060 XT continued to pull about 178W. Its efficiency is better than that of the 9070 XT here, which is beyond the peak point for efficiency for the V-F curve for the 9070 XT. The 9060 XT is lower in framerate, but its lower power draw allows it this efficiency lead over its larger alternative.

The 5060 Ti ends up slightly more efficient than the 9060 XT, with both pulling around the same power and pushing a framerate within single digits of each other. This is a good showing for AMD’s new card.

Efficiency: FFXIV 1440p

We’ll look at a worse one for AMD as well. Final Fantasy XIV has poor overall gaming performance for the 9060 XT, which means the 9060 XT ends up lower down in the chart. It’s at 0.51 FPS/W, which at least has it improved on the 9070 XT and 7800 XT. The card is also tied with Intel’s B580, but related to NVIDIA, it’s behind the RTX 4060 and significantly behind the 5060 Ti.

RX 9060 XT Conclusion

8GB cards can exist, but we don’t think they should at $300. If AMD wants to sell an OEM piece of sh*t edition, the company is welcome to do that, but we think those should be relegated to true low-end devices. 

Selling the lower VRAM alternative opens up the door to less informed users, who are likely not in our audience, and getting them tricked into buying something worse than they’re led to believe. It also artificially deflates the MSRP when people talk about pricing.

As far as performance versus the RTX 5060, it’s at least better in nearly all instances. It’s often around levels of the 3070 Ti. It tends to be between the 3070 and 3070 Ti in a lot of tests. 

The 16GB 5060 Ti is currently $490 without a combo and without open box. The 5060 Ti is up to 14% better in our 1080p tests, up to 17% better at 1440p, and up to 22% better in our 4K tests. In ray-tracing games like Black Myth, it really pulls away, but that depends on the game.

As for price commentary, we’re not going to delve into that yet until we can see where the prices land. We’ll then do a GPU pricing comparison once things settle down as usual.