We provide the rocky history of WinRing 0 and share how important it’s been to the PC industry

The Highlights

• The WinRing 0 driver has propped-up the fan control and RGB industry for over a decade
• Security updates have broken WinRing 0’s code
• This has resulted in a mad dash for hardware vendors to re-develop RGB and fan control software

Table of Contents

• AutoTOC

Intro

There’s one piece of obscure 2007 code that has propped-up the entirety of the fan control and large parts of the RGB industry. It’s a driver called WinRing0, and now, the software that uses it to control hardware through Windows is breaking.

If you're like us, you've always felt some sense of looming dread when you install a piece of software that controls fans or LEDs.

There's something ominous about a half-baked app with a broken UI taking control of hardware in meat space.

Editor's note: This was originally published on April 17, 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

Writing, Research

Patrick Lathan

Video Editing

Vitalii Makhnovets
Tim Phetdara

Writing, Web Editing

Jimmy Thang

As always, being pessimistic about this kind of thing has eventually paid off, with Hyte emailing us in March and The Verge posting a story about WinRing0 being flagged as a threat by Windows Defender (that article is worth a read for the statements provided by several developers). 

We contacted our own list of developers, and then reached out to Wendell from Level1Techs to help us talk through the technical aspects. This article explores the history and story of the WinRing 0 driver.

History Part 1: WinRing0's Creation

WinRing0 is a library originally released in 2007 by Noriyuki Miyazaki [宮崎 典行] (AKA hiyohiyo), and he regrets it.

The developer is best-known for CrystalDiskMark and CrystalDiskInfo. According to the active GitHub repository, "WinRing0 is a hardware access library for Windows" and "WinRing0 library allows x86/x64 Windows applications to access I/O port, MSR (Model-Specific Register), [and] PCI." 

Basically, WinRing0, the driver, is a unique open-source window into hardware. Over the years, it's become the equivalent of that XKCD comic for small developers who can't afford to develop and certify their own loopholes for controlling hardware like RGB LEDs and fans. 

If you're part of a small team that wants to distribute software for monitoring or controlling any of the hardware in a PC, WinRing0 has been the go-to option.

Hiyohiyo announced the end of development in February 2010, stating (in Japanese) that "WinRing0 is essentially a library that should not exist [...] I wanted to share the joy of low-level programming with as many developers as possible, so I developed and released WinRing0 after fully understanding the various issues, but I had no choice but to accept that this is no longer acceptable in today's age." 

He repeated that sentiment to us in an email, saying that "I consider it a complete youthful indiscretion on my part not to have accepted the changing times." 

The final update from hiyohiyo was WinRing0 2.0.0 in July 2010, where he intentionally removed almost all functionality, apologized again, and described the project (again, in Japanese) as a "big failure."

There’s something sad about that sentiment. For better or worse, the WinRing0 driver was actually not a big failure: A ton of hardware companies transacting hundreds of millions of dollars in revenue have relied upon it; however, this is probably why hiyohiyo views it as a failure.

Technical Explanation

Given what WinRing0 is -- a method of low-level access to hardware -- it makes sense that hiyohiyo has distanced himself from the project so much, especially since he currently collaborates with Microsoft. The releases of Windows Vista in 2007 and Windows 7 in 2009 made it increasingly clear that Microsoft is no longer in the business of letting you f*ck around with this stuff: 

Windows was moving away from low-level programming. The idea of old-school unrestricted memory access is scandalous these days. As Martin Malik of HWiNFO stated to The Verge, "since the driver has access and doesn’t restrict the range, it can read/change other processes, secrets in memory or protected kernel registers. This is very dangerous." As hiyohiyo stated when closing WinRing0 development 15 years ago, "If you think about why the OS restricts access to I/O ports, physical memory, MSR, etc., and why signing kernel-mode drivers is mandatory since Vista x64, you will understand."

We don't want to get too into the weeds here, but kernel-mode is the alternative to user-mode. 

We interviewed Wendell from Level1Techs, who went on to explain:

“What is the kernel? [You may have] heard of the Linux kernel but Windows has a kernel, too. So the kernel is responsible for management of your system; so process management, memory management, hardware abstraction, security isolation, and system calls, which is like a programmer's calls, like the kernel is going to provide this programmer's interface. You call [it] as a programmer and then the kernel goes off and does something. And so the buck stops with the kernel. So your programs just run and they don't have to deal with things like, ‘which processor am I running on,’ ‘how do I allocate memory?’ It just says I would like to allocate memory and the kernel [asks] how much memory would you like and then you get an address and then that's all handled; memory management, all of the abstraction for all those kinds of things. So the kernel is really the smallest, lowest part of your operating system and it is typically engineered to be as uncomplicated as possible. It's only as complex as necessary to do the task and if it has bugs that leads to a lot of problems, not just in terms of system instability but also security issues and that sort of thing.

Sometimes it's fun to think of it abstractly. Your computer is a bus and all of the apps on the bus are the passengers. The kernel is the driver of the bus and your computer hardware is like the engine, the wheels, the door, the brakes, and that kind of thing. The driver gets to decide how to use everything safely and effectively and if one of the passengers wet willies the driver then that's bad because it may put everybody in danger.”

The only reason that analogy is a bit confusing is because Wendell uses the word “driver” to explain the operation of the vehicle and he uses the word bus to explain the vehicle. 

With that in mind, let's take time to explain WinRing0's namesake: security Ring 0. Wendell elaborates, “There's a lot of ring 0 drivers as it turns out. Ring 0…kernel mode. I'm not a Windows developer [as my] day job but kernel operating system…it kind of makes sense. The things that are close to the hardware are ring 0 and they're supposed to have a relatively low surface area. If you are running an application and the application does something bad, which is ring 3, I believe, the application crashes. If you're running something at ring 0 and it crashes, it has the potential to affect the entire system and so the entire system will crash. Windows blue screens are probably ring 0. What has really accelerated Microsoft giving the boot to ring 0 is the CrowdStrike thing. This has been a problem forever but the CrowdStrike thing taking out the vast majority of infrastructure that runs Windows and Crowdstrike…Microsoft sees this as a problem and so this is basically a casualty of war.

Ideally you have things running in user mode ring three, all things running in user mode ring three. And so all of your software runs at ring 3 and the driver is very small and very low level and very lightweight and doesn't need to run quite as low level as ring 0 but is still sort of in the administrative permissions mode. But at a very low fundamental level, you can use software to update your BIOS and that is a pre-boot environment. You could have malware that lives in your BIOS. I would prefer having a motherboard that has a jumper so that when I want to re-flash the BIOS, I physically have to move a little switch to say yes.”

When we brought up Asus Armory Crate, Wendell added, “It goes the other way too, the BIOS could run arbitrary software.”

Kernel-mode drivers are almost always hardware device drivers, and within the x86 structure these (typically) occupy the highest security ring alongside the kernel: Ring 0. 

This is why a device driver literally named WinRing0 getting handed out to anyone who wants it might be a little alarming to Microsoft. As confusing as it is, Wendell's still pretty positive on the basic concept of security rings: “The ring 0, ring negative 1, ring 1…that’s all very tightly coupled with hardware features of x86 to provide isolation, which is great. There's different approaches from AMD and Intel, but there is there is a hardware aspect of this that is very nice for users as well so it's not just like you're entirely reliant on a 100% Microsoft software solution but a lot of this is how Microsoft has chosen to implement the various security levels but it dovetails with a lot of functionality that is at the hardware level, which is nice because the hardware is trying to protect you from code that shouldn't be executed.”

Digital Signatures

Microsoft's method for mitigating those concerns has been to require digital signatures for kernel-mode drivers in all Windows versions since 64-bit Vista. A digital signature is a certificate issued by a "trusted Certification Authority" (CA) that verifies that: "the file, or the collection of files, is signed. The signer is trusted. The certification authority that authenticated the signer is trusted. The collection of files was not altered after it was published." 

Back when WinRing0 was first published, individuals (in Japan) could sign drivers themselves, which hiyohiyo did. More expensive and difficult-to-obtain Extended Validation (EV) certificates were required starting in the Windows 8 era, and they're only issued to businesses, but old drivers were grandfathered in.

Over the years when installing a piece of software, you might have seen some kind of popup about the driver signatory, the lack or the presence of a signature. And we see this a lot with the prototype versions of software where they haven't signed it yet but as for why digital signatures are a useful idea in general, we turn again to Wendell who stated:

“As part of Microsoft’s strategy to deal with…driver signing, in general, any kind of executable signing is actually sort of fun and interesting. It’s a fun and interesting way of approaching security. If you right click on basically any executable on any modern Windows system and you look at the properties, you can see that the executable is digitally signed. That's an identity thing [that indicates] this is from [a particular] company. Drivers are a great way to hide malware and so it has to kind of be a walled garden and so the certificates you have on a website are really not [too] different or the executables from programs are really not [too] different from what you have for a driver. Basically you create the driver. You submit it to Microsoft and well, the submit-it-to-Microsoft process doesn't actually technically have anything to do with signing, but theoretically, Microsoft looks at you as a company and says ‘Okay, yes, we're going to be able to do business with you.’ And you get something that you can sign that is trusted and it is it is the standard certificate signing process where [you say] ‘here is my certificate’ [and] I'm going to send this somewhere that will then say: ‘okay, yes, we are going to sign the certificate that you have asked for except instead of being based on a hash or something ephemeral, it's based on the hash of the actual binary of the driver.’ And so this driver with this hash has been signed and if somebody tampers with the driver or changes it then the cryptographic signature will no longer match and the driver doesn't work anymore and so it's a nice way to affirm that something has signed off on the contents of this driver and this driver is good.”

Wendell also interestingly pointed out that CAs can be broken into and certificates can (and have been) stolen, but that's a subject for a different time.

So, hiyohiyo apologized for pulling the plug and refusing to maintain WinRing0's certification back in 2010, seemingly with the expectation that its certification would be pulled and everyone's projects would break: 

"WinRing0 was discontinued without any alternative plan in order to avoid the worst case scenario of the signature being revoked" and  "if the digital signature for WinRing0 is revoked, all WinRing0-based applications will be unable to start in an x64 environment."

History Part 2: WinRing0's Adoption

That brings us to the second part of WinRing0’s history.

WinRing0 actually became a foundational element of many, many projects, and some of those projects—like Open Hardware Monitor, later forked as LibreHardwareMonitor—would themselves become foundational to even more software on top of that. So there are nested layers of reliance on something that hasn’t really been even maintained or even liked by its original developer for 15 years. 

Seriously: You have very likely encountered WinRing0 in some capacity, and with the changes Microsoft is making for security reasons, a lot of those software encounters would no longer work today.

And that’s for good reason: Over the years, hiyohiyo's concerns were repeatedly validated.

In 2019, HP got in hot water for including WinRing0 pre-installed in its HP Touchpoint Analytics service "preinstalled on most HP PCs." This became a massive security concern from one of the biggest OEMs.

In 2020, WinRing0 was named in another CVE, or Common Vulnerability and Exposure, for EVGA's Precision X1. In 2021, it was Crucial's turn. Even though specific software was called out each time this happened, HP, EVGA, and Crucial were using the same 1.2.0 version of WinRing0 that everyone else was. 

As GermanAizek put it to us, "The driver was made in 2007. CVE in 2020. Microsoft started blocking it in 2025. Vulnerability has been around for 18 years." As for why Microsoft hasn't blocked it before now, according to OCCT, "They haven't done it yet because big corporations were lazy enough to use it in their software in the past, so that would invalidate their own software, so they cannot do it right away."

And the list of software that has used it at some point, and therefore software that has had vulnerabilities and attack vectors, is huge: CapFrameX (but not PresentMon), Precision X1, Crucial MOD, HP Touchpoint Analytics, SignalRGB, OpenRGB, and many more are on the list.

The issue isn't that Precision X1 or Crucial MOD or any of the vast array of affected software (CapFrameX, OpenRGB, SignalRGB, at least some versions of Afterburner, et cetera) are compromised: the issue is that they install an insecure driver (WinRing0) that's then accessible to any other software that wants it, including malware. 

This is precisely what happened with actual malware SteelFox starting in 2023; the vulnerability is real and has been actively exploited for profit. This isn’t just some proof of concept, this is an actual, in-the-wild malware that has been used to illicitly make money. 

Calling it "theoretical," as CapFrameX did, is irresponsible and dangerous, and it's not really relevant whether the software that installs the driver is itself safe. To quote OCCT:

"It's vulnerable as f*ck."

And here’s what Wendell thought, “If you say the last time the driver was meaningfully updated was in 2008 and it has not yet been exploited by malware, then that's a miracle.” We had to interject and say that it has been exploited by malware. 

For another example of a Ring 0 driver problem (not WinRing 0), check out what Wendell had to say about Crowdstrike, “So what happened was Crowdstrike has a ring 0 malware detection driver and Crowdstrike is otherwise very good software. It's very effective at what it does. It's an interesting security architecture. They made a mistake in their software and as a result of the mistake, the system tried to jump to memory address zero or start executing memory address zero. I don't really remember exactly what the details were but it was something obviously incredibly stupid and there was no safety rails for anything at this level and so systems would crash. And it was an impossible situation because the system would [consistently] boot and crash. If you were lucky after the 20th or 30th time, it would do that, the system would notice and deal with it and so Microsoft is saying ‘this is the wild west. We’ve got to deal with this ring 0 problem immediately and software like CrowdStrike cannot run at ring 0. We as operating system vendors have to provide a lower level facility to let these software vendors do what they need to do but without compromising the integrity of an update process without compromising the integrity of a boot process to provide fallbacks' and that sort of thing. As a result of that…I mean, internally, Microsoft has known this is an issue almost since day one. They didn't care until millions of machines had very large problems, basically every crowd customer that got the update.”

Beyond the current wave of Windows Defender alerts, WinRing0 and similar drivers also have a tendency to get flagged by software like Easy Anti-Cheat due to their ability to read and rewrite memory. You can make your own judgement about how serious the issue is, but these are not false positives. We want to make sure that’s clear. It isn’t a “false positive,” it’s just a true positive.

As hiyohiyo stated fifteen years ago: "although a general-purpose hardware access library such as WinRing0 1.x is very useful for prototyping, developers would need to develop dedicated device drivers for public release." 

But there needs to be a better, secure solution to gain access to this control and hardware. There is one and there has been one. As a developer, the 100% proper by-the-books response to this (from talking to numerous people) is to drop WinRing0, develop your own dedicated driver for your specific product, and obtain a signature for it. 

This is apparently the path that EVGA took back in 2020 after that CVE we mentioned.  

New signatures for kernel-mode drivers are really only accessible to large companies, though, with smaller dev teams unable to afford dedicating their time and money (in recurring payments) to the process, not to mention the software development work. 

Other manufacturers, including Hyte, have informed us that EVGA was somewhat propping-up fan control and RGB software by getting signatures on the driver. We’ve had a tough time trying to verify some of these claims, but that seems to be the belief held by, for example, Hyte. 

Therefore, WinRing0 has been eternally recycled and eternally frozen at vulnerable version 1.2.0. 

If you dig around in LibreHardwareMonitor's source code (for example), it references WinRing0.sys 1.2.0.5 from July 2008, which makes sense: hiyohiyo's next release included a reference in a patch note, saying that "it would have been an easy fix if only a digital signature could be obtained, but since the kernel mode driver cannot be updated, this was scrapped." 

According to Martin Malik of HWINFO64, this day of reckoning has been a long time coming, with Microsoft repeatedly warning that the driver would be blocked.

Again, we've heard unconfirmed reports that EVGA possibly took up the maintenance for WinRing0's digital signature in the post-2010 era, possibly arranging for its renewal (as we understand that certificates expire over time) or just convincing Microsoft not to revoke it. If EVGA had any involvement, it probably ended in 2020 when the company stopped using WinRing0, or at least in 2022 when the company basically halted operation. Microsoft's statement to The Verge that "we are aware of reports about gaming and monitoring applications being flagged as a threat due to the use of unsigned versions of the WinRing0 driver" implies that the driver is now unsigned, which could be a further clue that EVGA was doing some kind of upkeep behind the scenes.

Somehow, we continue to learn EVGA’s impact beyond its GPUs.

The Future of WinRing0

All of this is a problem, because there are limited tools to control hardware through the OS -- and for good reasons -- but there needs to be something, and currently, many of those tools are breaking or broken. Or insecure.

That brings us to the future of WinRing0.

The easiest solution to all this would be to patch WinRing0 itself. After hiyohiyo's last constructive contribution in 2009, Herman Semenov [Герман Семёнов] (AKA GermanAizek) took over maintenance in 2019, initially with the goal of optimizing crypto mining with access to CPU MSR registers. As he stated to us, "around 2023, many people wanted to build WinRing0 Windows driver themselves to increase mining hashrate, even though it was much more difficult than just mining on Linux."

In a weird way then, crypto mining potentially provided something directly useful to those controlling hardware for non-mining use cases.

Development accelerated in 2023 as other members contributed to the project, adding x64 support and fixing some BSOD triggers in the old driver. Eventually, the team applied patches to address the open CVE from 2020. Critically, this fork of WinRing0 remained unsigned: only the un-optimized, insecure version from 2008 had the valid signature vital to projects like LibreHardwareMonitor.

This is where HYTE has stepped in. HYTE originally contacted us with the story, stating that it wants to take the version of WinRing0 that GermanAizek's team has been updating, submit it to Microsoft for signing, and fork LibreHardwareMonitor to integrate the patched, signed driver. HYTE would then take on the responsibility of paying Microsoft, basically replacing EVGA’s assumed role in this chain.

The direct benefit is that HYTE's own software can continue to function, while the rest of the industry gets to keep using WinRing0 (and LibreHardwareMonitor) without getting auto-quarantined by Windows Defender.

GermanAizek told us that "these fixes restrict the use of the driver only to programs running with administrator rights." This is certainly safer, but (as Martin Malik of HWiNFO warned The Verge), this just means that an app has to be run as admin before it can access the driver. 

We asked Wendell about this. Specifically about running things as admin and how much that might help. Here’s his response:

“That's probably not unreasonable. Short of Microsoft getting involved and offering a better solution or somebody that is that deep in the Microsoft kernel driver developer ecosystem, that's probably what it would take: somebody that has very deep intricate knowledge of the operating system and also knows what the operating system is capable of. As far as I know, you're on your own to implement a lot of the functionality that would be needed to do that. So this driver is probably still your best hope to do that. Microsoft probably doesn't want to adopt the driver, which would also be a reasonable outcome. At the same time, Microsoft probably doesn't want to re-implement the functionality that's in the driver, but how this is usually done is you peel away the minimum functionality and you stuff that in your ring 0 driver and then you have all of the other stuff live somewhere else. And that ring 0 driver, you trust not to be able to be manipulated to access memory, it's not supposed to or write to a bus address that it's not supposed to be able to.”

So, as Wendell helped us understand, the idea that a combination of patches and signatures can fix the root cause of the problem is arguably misguided. 

We contacted Franck Delattre of CPUID (CPU-Z, HWMonitor), who explained that CPUID has had similar difficulties with its own software. "In order to fix the problems, we had to move a big part of the user code into the kernel code, in the unique goal of reducing exposure. We could do that because only our code uses our driver, but for a generic driver like WinRing0, this was simply not possible since its functions were used in a different context by the different application. To go further, this means that no replacement of WinRing0 is possible, at least not with the same genericity that WinRing0 provided until today." 

In other words, the thing that makes WinRing0 uniquely useful is the same thing that makes it dangerous. 

GermanAizek is literally the frontman for the "fixed" version of WinRing0. 

He told us that "personally, I migrated to Linux and BSD systems because Windows has become really insecure, and as a Unix developer, such operating systems really seem convenient to me." 

He also openly asked that developers use the InpOut32 driver instead of WinRing0 (although we've seen other developers express concerns about that as well).

OCCT has also announced that it will be providing a publicly-available but closed-source alternative to WinRing0, and it's possible that other organizations will follow suit.

Wendell informed us that there are other bigger-picture alternatives, “For sensors and fan speed, one way that you could solve this architecturally is to just move it to a USB controller. That's slightly more cost or if somebody wants to build in a USB client interface then that's probably a marginal cost increase. I'm slightly surprised it hasn't gone in that direction but I'm also slightly surprised because this is a problem for Windows server in the context of the system management bus because servers need access to the system management bus and kind of hilariously, you have the out-of-band management that also has access to the system management bus so like servers have a whole other computer inside them that has access to the system management bus and the same controllers and so you can use that computer within a computer to monitor the sensors. You could just not have that and plug it into USB in the case of client computers. Like I say, we put important things on the system management bus and so like controlling CPU voltage probably should be on the system management bus. Controlling fan speed…You could probably do that through USB, but when it's through USB, the chipset and other things probably are not able to control fan speed. So you end up with a chipset that needs a system management bus so low-level parts of the system can make those controls but user overrides have to come through another path like through USB or something that's low security. Or Microsoft can provide a reasonable facility that is reasonably locked down to access the system management bus facility.”

Conclusion

That’s the story of how this small piece of code has supported an entire industry and its software for 15 years now, even in spite of its own developer disowning it and regarding it as not only a mistake, but a failure. We feel bad for hiyohiyo who now is powerless to stop people from using his youthful development project, but these multi-million and billion dollar companies have the resources to develop a responsible alternative. That includes Microsoft, Razer, and everyone else. 

That brings us around to what power an end user has, if any.

Our recommendation is to do what your antivirus software says: if Windows Defender quarantines WinRing0, let it happen, and if anyone tells you to ignore the warnings, treat them with extreme skepticism. Some manufacturers and developers have called these “false positives,” but they are not. 

They are real positives, and there are real vulnerabilities that have been used which can exploit your machine.

Microsoft appears to have paused the "ban" as of this writing, but it's only a matter of time. If everything goes according to plan, though, the patched driver should be usable soon thanks to HYTE, at which point you can decide whether requiring admin privilege for access meets your personal standard for security. 

For whatever it's worth, the Windows Dynamic Lighting RGB control feature continues to be developed, although it doesn't feel great to be railroaded into using it just because Microsoft bricked the alternatives. 

Still, it's probably the right direction for Microsoft with Wendell stating, “There is one aspect of this where Microsoft is doing the right thing and that is RGB control. Windows 11 allows you to control RGB directly in the operating system. Microsoft [shouldn’t take] half measures here and add some fan controls and or at least provide a programming interface. [Microsoft doesn’t] have to [provide] a GUI for fan control like it did with RGB control but wherever that's plumbed in, [Microsoft should] go ahead and plumb in the other stuff. It's really not any more complicated than that.”

Thanks to the various developers that provided quotes for this piece, as well as Wendell.

We provide a detailed breakdown and timeline of the tariffs impacting the United States

The Highlights

• The US has imposed wide-spread tariffs on most of the countries in the world
• Our timeline chronicles the shifting nature of the tariffs
• Tariffs have impacted the pricing and availability of hardware in the US

Our timeline below accompanies our deep-dive investigation into the impact of the tariffs roll-out on the computer hardware industry. Our investigation focused on not just the percentages, but specifically, the way in which they were announced and applied.
We flew around the country to meet with factories in the US, manufacturers who use contract manufacturing in other countries (like China, Vietnam, Thailand, Taiwan, Singapore, Japan, so forth), German manufacturers, and more. Our interview list included Hyte, Thermal Grizzly, Protocase/45 Drives, iBUYPOWER, CyberPower PC, Corsair Components, freight forwarding company Straight Forwarding, and Louis Rossmann of Rossmann Repair.
You can find that video embedded below. Below that, we have included the entire timeline of tariff changes that we could find. It is dated back to 2018. Please note that we have likely missed a few things, as this situation changes nearly daily and our regular coverage spectrum has not traditionally included following these changes; however, we believe we have compiled all of the major changes that are directly relevant to this story, particularly as it relates to computer hardware. This was a huge team effort at GN and required a massive investment in travel and time to complete our 3-hour documentary. If you find it educational, we ask that you please support us directly by buying something from our store. You can find high heat resistant project and soldering mats, PC building Modmats, T-shirts with paper launch GPUs or Honey Pots from our lawsuit against PayPal, copper-plated stainless steel mule mugs with thermal conductivity written on them, tabletop gaming dice with embedded inductors or cats, and more. Thank you.

Credits

Host, Writing, Lead Editing

Steve Burke

Lead Camera, Storyboarding, Editing

Vitalii Makhnovets

Pre-Cut, Labeling

Mike Gaglione
Tim Phetdara

Animations, Labeling

Andrew Coleman

Writing, Research

Jimmy Thang
Ben Benson

US Tariffs Timeline

This timeline is sorted chronologically. We have attempted to present it as neutrally as feasible and from a place of covering the impact and events. As a part of this effort, we are also including links to sources with ideologies we may not agree with, but which we believe are appropriate for establishing the timeline of events.

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 a government announcement happens during briefings or interviews with reporters.

Background: Tariffs During The First Trump Administration & Biden Administration 

• Jan. 22, 2018: The Trump Administration imposed tariffs on solar panel components and residential washing machines. Solar panel components had a 30% tariff, which would decline over four years. The administration levied a tariff of 20% on the first 1.2 million imported washers, which could increase to 50%. 
  • Government Trade Release 
  • NPR
• March 8, 2018: President Trump announced a 25% tariff on steel imports and 10% tariff on aluminum imports for many countries, not including Canada and Mexico. 
  • Federal Register (1, 2)
  • NPR 
• May 31, 2018: The Trump Administration imposed a 25% tariff on steel imports and 10% tariff on aluminum imports for Canada, Mexico and the European Union. The administration said it had “reached arrangements on steel with Australia, Argentina, and Brazil, and with Australia and Argentina on aluminum.”
  • U.S. Government Document
  • Washington Post 
• June 29, 2018: Canada announced tariffs targeted at U.S. steel and aluminum products, in addition to other goods. The tariffs would go into effect on July 1, 2018. 
  • Canadian Government Release 
  • Quartz 
• Sept. 18, 2018: The United States issued $200 billion in tariffs on Chinese imports, citing the need to protect “intellectual property.” 
  • Government Trade Release
• May 17, 2019: The United States removed tariffs for steel and aluminum imports from Canada and Mexico.
  • Government Trade Release 
  • CNBC
• July 1, 2020: The USMCA - a trade agreement among the United States, Canada and Mexico - goes into effect. 
  • USMCA Text

• Sept. 30, 2021: In a Politico interview, a Biden administration official said the United States will “build on” existing tariffs that are levied against China.
  • Politico
• May 14, 2024: The Biden Administration increased tariffs on imports from China for electric vehicles, solar cells, lithium-ion vehicle batteries, and semiconductors, among other products. Tariff increases ranged from 25% to 100%. 
  • Government Trade Release 
  • U.S. Department of Commerce Fact Sheet
• Sept. 13, 2024: The Biden Administration finalized tariff rates and exceptions from the May tariff announcement. Updates included new timing, rates and exclusions for some medical devices, solar manufacturing equipment, and wafers. 
  • Government Trade Release 

Tariffs During The Second Trump Administration 

Monday, Jan. 20 

On inauguration day, President Trump said the U.S. government would implement 25% tariffs on imports from Canada and Mexico, starting on Feb. 1. 

• Sources
  • Associated Press

Sunday, Jan. 26 

The Trump Administration said it would impose 25% tariffs on imports from Colombia after an immigration dispute. In response, the Colombian government threatened 25% tariffs on imports from the United States. The situation deescalated after the two countries reached an agreement.  

• Sources
  • Truth Social post
  • Axios
  • Associated Press

Saturday, Feb. 1

The Trump Administration announced 25% tariffs on imports from Canada and Mexico. Energy imports from Canada would have a lower tariff rate of 10%.

The Trump Administration announced 10% tariffs on imports from China. 

The tariffs on Canada, Mexico, and China would go into effect on February 4. 

• Sources
  • Executive Order (Canada)
  • Executive Order (Mexico)
  • Executive Order (China)
  • White House Fact Sheet
  • NPR

Monday, Feb. 3

The Trump Administration agreed to a 30-day pause on tariffs with Canada and Mexico after negotiations. 

President Trump told reporters that the European Union could face tariffs soon.

• Sources
  • Executive Order (Canada)
  • Executive Order (Mexico)
  • Justin Trudeau post (Canada) 
  • Claudia Sheinbaum Pardo post (Mexico) 
  • NBC
  • BBC

Tuesday, Feb. 4

The United States imposed a 10% tariff on imports from China, as announced on February 1. 

China announced new tariffs on imports from the United States, including a 15% duty on coal and liquefied natural gas. China also announced 10% tariffs on imported crude oil, agricultural machinery, and pickup trucks. These tariffs would go into effect on February 10. 

China announced export controls on several minerals and began investigating Google for antitrust violations. 

• Sources
  • China - Government Documents (1, 2, 3)
  • Associated Press
  • CNN
  • BBC

Monday, Feb. 10

In a continuation of tariffs from 2018, the Trump Administration announced a 25% tariff on imported steel and aluminum, which would go into effect on March 12. The administration stopped providing exemptions for steel and aluminum imports from Argentina, Australia, Brazil, Canada, Japan, Mexico, South Korea, the European Union, Ukraine, and the United Kingdom. 

To avoid confusion with other tariff names, we will refer to these as steel tariffs and aluminum tariffs going forward. 

China’s tariffs on imports against the United States went into effect, as announced on Feb. 4. 

• Sources
  • Executive Order
  • White House Fact Sheet
  • NPR (1, 2)

Thursday, February 13

The Trump Administration announced the Fair and Reciprocal Plan, a trade plan intended “to reduce [the United States’] large and persistent annual trade deficit in goods and to address other unfair and unbalanced aspects of [its] trade with foreign trading partners.” 

The White House said the trade plan would determine “the equivalent of a reciprocal tariff with respect to each foreign trading partner.” 

To avoid confusion with other tariff names, we will use the reciprocal tariff naming for these specific tariffs going forward.  

• Sources
  • White House Memo
  • White House Fact Sheet
  • Associated Press

Monday, March 3

The Trump Administration raised tariffs on imports from China from 10% to 20%. 

• Sources
  • Executive Order
  • Associated Press

Tuesday, March 4

The United States’ 20% tariffs on imports from China go into effect. 

The United States’ 25% tariffs on imports from Canada go into effect. The United States taxed Canadian energy imports at 10%, rather than 25%. 

The United States’ 25% tariffs on imports from Mexico go into effect. 

• Sources
  • White House Fact Sheet
  • Axios

China announced tariffs of 10%-15% on U.S. farm and food exports and added 10 U.S. companies to its unreliable entity list. 

• Sources
  • China - Government Documents (1, 2) 
  • Reuters (1, 2)
  • Xinhua

Canada announced 25% tariffs on many U.S. imports. The Canadian government said the tariffs would cover U.S. goods worth C$30 billion (~$20.7 billion) immediately. 

• Sources
  • Canadian Government Release
  • Justin Trudeau post 
  • Associated Press

Wednesday, March 5 

The Trump Administration paused tariffs on cars coming into the United States from Canada and Mexico for one month. According to White House officials, the pause happened after President Trump spoke with leaders of Ford, General Motors, and Stellantis. 

• Sources
  • Associated Press
  • CNN

Thursday, March 6

The Trump Administration paused tariffs on some products from Canada and Mexico that are compliant with the USMCA (United States-Mexico-Canada Agreement), a free trade agreement. 

White House officials informed media outlets that 50% of Mexican imports and 38% of Canadian imports are compliant with the USMCA.  

• Sources
  • Executive Order (Canada)
  • Executive Order (Mexico)
  • White House Fact Sheet
  • ABC News
  • CNBC

Monday, March 10  

As announced on March 4, China began imposing 10-15% tariffs on many farm products from the United States, including soybeans, wheat, corn, beef and chicken. 

• Sources
  • Associated Press
  • China - Government Document 

Wednesday, March 12 

The United States’ 25% tariffs on steel and aluminum imports go into effect. 

Canada increased its tariffs on U.S. imported steel and aluminum products by 25%. The Canadian government said the affected products include tools, computers, servers, and display monitors, among other products. The tariffs would go into effect on March 13. 

The European Union responded with its own tariffs against the United States. The EU said that its countermeasures could “apply to US goods exports worth up to €26 billion.”

• Sources
  • Canada - Government Document Release
  • EU Release
  • EU Fact Sheet
  • CNN
  • Associated Press
  • Axios 

Thursday, March 13

President Trump threatened the European Union with a 200% tariff on European alcohol. 

• Sources
  • Truth Social 
  • NPR

Monday, March 24

The Trump Administration issued an executive order that would impose a tariff of 25% on imports from countries that import oil from Venezuela, starting on April 2. 

• Sources
  • Executive Order
  • Associated Press
  • Axios

Wednesday, March 26

The Trump Administration announced a plan to impose 25% tariffs on imported automobiles and car parts, such as engines, transmissions, and electrical components. Tariffs would begin on April 3. The administration included exemptions for USMCA-compliant auto parts.

• Sources
  • Executive Order
  • White House Fact Sheet
  • NPR
  • CNN

Thursday, March 27

Industry impacts:

• Hyte announced higher prices due to incoming tariffs: Hyte blog 

Wednesday, April 2 

Consistent with the Fair and Reciprocal Plan from February 13, the Trump Administration announced reciprocal tariffs on many countries. In addition, the administration announced a 10% tariff on nearly all countries. 

The administration said the 10% tariff rate would go into effect on April 5. The reciprocal tariffs would go into effect on April 9. 

As stated on February 13, the administration implemented different reciprocal tariff rates for many trading partners. Examples: 

• Vietnam - 46%
• Thailand - 36%
• Taiwan - 32%
• Japan - 24%
• European Union - 20% 

The administration added a 34% reciprocal tariff to China, on top of the existing 20%, bringing the total to 54%.

• Sources
  • Executive Order
  • Executive Order (China)
  • White House Fact Sheet
  • Federal Reciprocal Tariff Rates (Annex I)
  • Presidential Tariff Actions
  • CNN
  • NPR

Thursday, April 3

As announced on March 26, the Trump Administration's 25% tariffs on imported automobiles went into effect. The administration delayed tariffs on automotive parts until a date no later than May 3. 

• Sources
  • Federal Register 
  • CNN 

Industry impacts:

• META PCs issued a statement about possible price increases: Statement

Friday, April 4

China’s government said it will impose additional tariffs of 34% on U.S. goods starting on April 10. China announced export controls on several rare earth minerals. 

• Sources
  • China - Government Documents (1, 2)
  • CNN
  • Tom’s Hardware

Industry impacts:

• Hyte published an update on price increases: Hyte statement. 
• iBUYPOWER issued a statement about the tariff impacts: iBUYPOWER statement.
• Nintendo delayed Switch 2 pre-orders to the United States: Statement to IGN.

Saturday, April 5

The Trump Administration’s 10% tariff goes into effect for many countries, as announced on April 2. 

• Sources
  • Federal Register
  • Reuters

Monday, April 7

Industry impacts:

• Framework paused sales of several Framework Laptop 13 models to the United States: Twitter statement. 

Tuesday, April 8

The Trump Administration increased the tariff rate on imports from China by 50%. Through an executive order, the administration amended the reciprocal tariff rate for China from 34% to 84%. This tariff rate increase was in addition to the 20% tariff rate from March 4. 

• Sources
  • Executive Order
  • CNN

Industry impacts:

Reports surfaced of Razer stopping laptop sales to the United States. Razer did not publicly confirm this, but U.S. consumers were unable to buy laptops on Razer’s website. 

• Sources
  • Tom’s Hardware
  • The Verge

Wednesday, April 9

The Trump Administration’s reciprocal tariff rates went into effect shortly after midnight. 

Later that day, with the exception of China, the Trump Administration paused many reciprocal tariffs for 90 days. The administration did not pause the 10% tariff on imports from April 5. 

China responded with an additional 50% tariff on U.S. imports. U.S. goods shipped to China now faced an 84% tariff, when accounting for the existing 34% tariff from April 4. 

The Trump Administration raised tariffs on imports from China to 125%.

The European Union approved new tariffs against the United States. The EU said details would come later. 

• Sources
  • China - Government Document
  • Executive Order
  • CNN
  • Associated Press 
  • CNBC
  • NPR 

Industry impacts:

• Per a report in Nikkei, a newspaper in Asia, Apple requested suppliers to air ship premium devices to the United States before the April 9 tariff deadline. 
  • Nikkei, MacRumors 
• Framework increased prices and delayed some shipments to the United States: Framework blog 

Thursday, April 10 

The Trump Administration clarified to reporters that the 125% tariff on Chinese imports was in addition to the previously announced 20%. The White House confirmed the tariff rate for imports from China is 145%. 

The European Union paused its tariffs against the United States. 

As announced on April 4 and April 9, China’s tariffs on imports from the United States go into effect.

• Sources
  • European Union Statement
  • The New York Times
  • CNBC 

Friday, April 11 

China increased tariffs on imports from the United States from 84% to 125%. China said the new tariff rate would go into effect on April 12. 

• Sources
  • China - Government Document
  • CBS

Industry impacts:

• Per a report in Commercial Times, multiple OEMs paused laptop shipments to the United States. The reported companies included Dell, HP, and Lenovo. Mashable reached out to the three OEMs for comment; Mashable has not published a comment yet. 
  • Commercial Times, Mashable 

Friday, April 11 - Saturday, April 12 

On April 11, the US federal government published guidance stating that some electronics, including smartphones and computers, are exempted from the reciprocal tariffs. Media reported that the exemptions are intended to help U.S. technology companies like Apple. 

• Sources
  • U.S. Customs and Border Protection Document
  • Executive Order
  • CNBC

Sunday, April 13 

President Trump said the United States will announce tariffs on semiconductors soon. An administration official clarified that tariff exemptions for electronics are “temporary.” 

• Sources
  • Truth Social
  • Associated Press
  • Axios
  • CNN
  • USA Today 

Monday, April 14

In a meeting with reporters, President Trump said he may give tariff exemptions to automobile manufacturers. 

• Sources
  • Associated Press

Industry impacts:

• NVIDIA announced that it will begin building AI supercomputers within the United States: NVIDIA blog

Tuesday, April 15

In a White House fact sheet, the administration said imports from China face tariffs of up to 245% on certain goods, including electric vehicles, when accounting for all previous tariffs. 

• Sources
  • White House Fact Sheet

Friday, April 18

The federal government announced fees for ships made in China destined for US ports. The fees could begin in October and increase over time. 

• Sources
  • Federal Register Notice
  • MSN (Reuters)

Sunday, April 20

The Chinese government warned countries about cooperating with the United States on trade at the expense of China’s interest. According to a translation via the BBC, the Chinese Commerce Ministry said, “China firmly opposes any party reaching a deal at the expense of China's interests. If this happens, China will never accept it and will resolutely take countermeasures.”

• Sources
  • China - Government Document
  • BBC
  • CNBC

Tuesday, April 22

The US Department of Commerce “determined that imports of solar cells from Cambodia, Malaysia, Thailand, and Vietnam are being dumped into the U.S. market.” The Internal Trade Commission, a separate agency, will make the final decision on tariffs. 

• Sources
  • U.S. Press Release
  • CNN

Helpful Timelines

• Associated Press
• The Guardian
• New York Times
• Peterson Institute for International Economics

Please consider watching our investigation for more information.

CPU benchmarks & comparisons across multiple generations for gaming and productivity tasks

The Highlights

• This data has been collected from our CPU reviews and benchmarks
• The data includes important caveats and disclosures relating to vetting processes for long-term support charts
• You can more easily determine if your CPU is in the list with Ctrl+F. The table at the bottom lists all CPUs detailed in at least one chart.
• This is a large, ongoing effort and will get updates at this URL permanently
• Please consider supporting this effort on our store

Table of Contents

• AutoTOC

Intro

This article is an entry in our GN Mega Charts series. All Mega Charts are listed on the Features page, including these:

• CPU Power Consumption & Efficiency Mega Charts
• CPU Cooler Mega Charts

This section contains disclaimers, limitations of the process, and disclosures relating to data quality control. We think that this is all important for your understanding of how this page works and so that you can adjust your own expectations and potential reliance on the data to calibrate with the two groups (“Active” and “LTS”); however, if you’d like to just jump straight to the charts and ignore all of that, you may bypass the wall of text and auto-scroll down with this link.

This article contains our ‘Mega Charts’ for CPU performance benchmarks, including our production tests (commonly referred to as “creation” benchmarks) and gaming tests. Our power testing can be found on the above-linked page and is isolated, as it tends to be more static.

This page will be regularly updated with the latest of our CPU benchmark performance numbers. It will consist of two types of charts: Long-Term Support (“LTS”) and Active. The long-term support charts have several special caveats, but are intended to be available to help people better determine upgrade paths. The LTS charts are more likely to contain older CPU results.

The page also includes links to CPU reviews and comparisons, such as historical AMD vs. Intel benchmarks. It will be updated on a slower cadence from our latest reviews (so you should always defer to those for the most recent numbers), but will be updated a few times a year with larger charts than are found in our reviews. This is for a few reasons, but one is that we shorten review charts due to video height limitations (16:9 aspect ratio). The other is that it’s just too crowded for the regular updates.

This page is intended to be used long-term for our Mega Charts. You can bookmark this page, as our future updates for CPU Mega Charts will land at this same URL. The update log will be posted at the bottom of the page so that you always know the latest data set. It will be updated a couple times a year, with more frequency updates in the CPU reviews themselves.

Credits

Test Lead, Host, Writing

Steve Burke

Testing

Patrick Lathan
Mike Gaglione

How to Use This

Even as data ages, it is often still relevant for comparison -- particularly for older CPUs which mostly stop receiving performance-affecting changes, such as microcode updates or Windows patches. We are constantly re-running our CPU tests to keep data fresh, but unfortunately, this refresh cycle means that it is difficult to stack more CPUs on the charts before some sort of major change comes in. For example, major Windows updates necessitate full re-tests for reviews, but may not be as important for someone who just wants to see their older CPU represented for a “good enough” gauge of where things fall.

That’s why we split these into the LTS (Long-Term Support) and Active charts. It allows us to maintain one older dataset that has more CPUs represented, at the cost of reduced insights gained from our most modern test methods. Active gives you that for more of a modern head-to-head. It’s the difference between precision (Active) and quantity of CPUs (LTS). This is the best balance that a small team can produce, especially since we provide this website completely ad-free (you can support us on Patreon or by buying something useful for your PC builds on our store).

Current Best CPUs (Generalized Recommendations)

The below is a simple list of CPUs that, at the time of writing (dated in the columns below), we think make sense or would make sense with caveats noted.

Pay careful attention to the second column. We may only recommend some parts under certain pricing conditions. Generally speaking, we do not recommend buying CPUs above MSRP. They come down pretty regularly, especially with launch cycles. For instance, when we first posted this, the 7800X3D was about $250 overpriced. It's still on the list so that people are aware of it, but we advise waiting for it to approach that MSRP marker or to be replaced with the pending 9000X3D parts.

For a better and more thorough list of Best CPUs, please check our Best CPUs of 2023 article. We will update for 2024 also.

Is Your CPU Missing From This Data? Here’s What to Do

We frequently receive questions from people asking where their particular CPU would land on a chart. There are thousands of CPUs that could be tested, so we obviously don’t have all of those listed. The best bet is to approximate the positioning by just thinking through the data that is present and using deductive reasoning. Newer builders may not realize that it is often this simple, so we’ll outline some concepts so that you can at least get a rough idea of where your part might fall. You can apply this to any reviewer’s charts. And of course, if we’re missing a part, there are plenty of qualified reviewers out there who may have something we’re missing (and we likewise try to cover what they miss) -- that’s the value of multiple qualified reviewers.

CPU Performance Interpolation/Deductive Reasoning Examples

Example 1: The Ryzen 5 1600AF is not present on the charts.

Solution: The Ryzen 5 1600AF is functionally an R5 2600, just at slightly different clocks. Looking up original reviews would get you this information, which you can then apply to modern charts. Looking at the R5 2600 in a chart is close enough to the R5 1600AF that you could base your decisions off of that part.

Example 2: The Ryzen 5 2600 is also missing from the charts.

Solution: Pull up a few old/original reviews of the R5 2600 and identify parts that are nearby or adjacent in performance. Look at a few charts, as some games can differ. Once you have found the most commonly comparable part, you can use that as a rough gauge on charts.

Example 3: The i7-10700K is missing from the charts.

Solution: If the i5-10600K and i9-10900K are present, it’s reasonable to assume that the 10700K is between them. Although this can have a relatively wide range, the reality is that, especially upgrading from something older, it won’t matter enough to hurt decision making on new processor purchases (since anything will be a huge upgrade).

Example 4: The Sandy Bridge i7-2600K is present, but not the i7-2700K

Solution: In situations such as these, where the part is basically just a slight change (example: 2700X vs. 2700, 2600X vs. 2600, 7700X vs 7700), you can just look at the one that is present and assume close enough performance to compare. This is again where it’s important to keep perspective: If the goal is to upgrade, being 2-5% off on the estimate isn’t going to meaningfully impact decisions if the alternative is no good modern data as reviewers move on.

As a last-ditch solution: You can look at games or benchmarks which are least likely suspected to have had major patches, such as GTA V, and calculate the percent difference between the target part and mutually present part on both an older chart and the new one. Then apply this to the modern chart to approximate or interpolate the rank. For example, if you pull up a chart from two years ago with the 8700K and 12700K on it and calculate the difference (typically, (new - old) / old), you can then apply that on the same game chart from modern times. This is the least perfect method because newer games may have architecturally evolved and may not be linear and older games get patches. You’d frankly be better off finding it somewhere else on the internet, but if you really can’t, this is a method that helps get at least something to work with.

DISCLAIMER: Data Accuracy Standards & Reduced Vetting Practices

For standalone reviews that receive full video treatment, we run through a quality control process that is intensive and often takes several days to complete. These are time-intensive, cost-intensive, and critical to the accuracy of our mainline reviews. We would never skip steps for the fully produced reviews.

But we want to publish more of the data we collect outside of the reviews process because it’d help people with purchasing decisions. We have been resistant to publishing the extra data because the ‘extra’ data doesn’t go through the same validation processes. It’s still useful and rarely has issues; however, our QC standards are high and we are careful with what we release. Anything we’re uncertain of or haven’t vetted fully is withheld until it clears those bars. 

But a lot of people would benefit from knowing where an AMD R7 1700 lands today, and outside of full revisits, we don’t have a good avenue to release data that is useful, likely up to our usual standards, but just not quadruple-checked. That’s what the Long-Term Support (LTS) charts are for. 

We’ve decided to release the extra data to try and cast a wide net to help people upgrading from older or more obscure parts, but are doing so with the upfront disclosure that the difference between the Active charts and the LTS charts is the validation process. In other words, data which exists on LTS charts but not Active charts should be understood by the audience as more likely to have some sort of data confidence issue or possible deviation from expected results. It is still unlikely, but more likely than Active charts. With that transparency and understanding, here’s the difference in our processes:

Active Chart Vetting Process

One of our biggest hangups for publishing a full list of our “Mega Charts” for CPUs has been that we simply cannot sustain the intensive QC process for a secondary dataset that contains CPUs that haven’t been published on the channel recently. We often still collect data for older CPUs, but may not publish it for time reasons (meaning that it was collected for internal review, but not fully vetted for publication). 

Here’s the full process, with some specialty confidential steps removed:

• The Test Lead for the tests (typically Steve) establishes a strict SOP for the test suite to ensure data consistency. The Test Lead also determines which software is tested, performs benchmark analysis for standard deviation and consistency, and determines the testing methodology
• The Test Lead (typically Steve) establishes a Test Matrix containing all CPUs to be tested, which tests they get, and data exporting practices
• The Subject Matter Expert (Patrick) builds software in collaboration with Steve for the test suite and builds the operating system configurations, establishing a “pre-flight checklist” for testing
• The Technician who runs the tests (primarily Mike, with assistance from Patrick and Steve and occasionally Jeremy) will check the results for correct resolution, correct settings, captures that prove the test completed correctly, image quality, and general errors. The technician performs re-runs as needed
• The Subject Matter Expert performs a secondary quality control pass on data. Any data which looks suspect of a bad pass, which requires manual filtering & removal prior to retests, will be deleted and flagged for the technician to rerun
• The Technician performs re-tests and then re-evaluates the results
• The Subject Matter Expert reviews them again. They are typically resolved at this stage and occasionally rejected entirely to move forward
• A Writer & Technician (typically Jeremy) performs a full pass on CPU export names, the hierarchical stack of the data and whether the hierarchy makes sense, and identifies potential areas where additional data may be necessary to confidently come to conclusions, then passes it to Steve. If Steve authorizes any of the re-evaluations from the Writer & Technician, they go back to the lab again for testing.
• Steve performs final intensive review, including running a variety of formulae for data consistency and evaluating results against archival results for consistency of scaling
• The Writer & Technician double-checks the finals. Any disagreement with what Steve passed is voiced and revisited. Once both are in agreement that it makes sense (or the suspect data is removed until validated), the data is sent to publication
• The video editor, who is often technical enough to identify obvious oversights, performs a ‘passive’ QC while editing and flags any issues. If any are found, they go through Steve for review, analysis, and either sign-off as accurate or correction

You’ll notice that a big part of the process is passing the results through multiple team members -- typically 4-5 people look at it before publishing. This is slow, but important. 

Long-Term Support Chart Vetting Process

The Long-Term Support charts will contain data which is not in reviews, but was used as internal gauges for where parts should sensically land. As a result, this is the process it has followed since it has not previously been published:

• (Same) The Team Lead for the tests (typically Steve) establishes a strict SOP for the test suite to ensure data consistency. The Team Lead also determines which software is tested, performs benchmark analysis for standard deviation and consistency, and determines the testing methodology
• (Same) The Team Lead (typically Steve) establishes a Test Matrix containing all CPUs to be tested, which tests they get, and data exporting practices
• (Same) The Subject Matter Expert (Patrick) builds software in collaboration with Steve for the test suite and builds the operating system configurations, establishing a “pre-flight checklist” for testing
• (Same) The Technician who runs the tests (primarily Mike, with assistance from Patrick and Steve and occasionally Jeremy) will check the results for correct resolution, correct settings, captures that prove the test completed correctly, image quality, and general errors. The technician performs re-runs as needed
• (Removed) The Subject Matter Expert performs a secondary quality control pass on data. Any data which looks suspect of a bad pass, which requires manual filtering & removal prior to retests, will be deleted and flagged for the technician to rerun
• (Removed) The Technician performs re-tests and then re-evaluates the results
• (Removed) The Subject Matter Expert reviews them again. They are typically resolved at this stage and occasionally rejected entirely to move forward
• (Removed) A Writer & Technician (typically Jeremy) performs a full pass on CPU export names, the hierarchical stack of the data and whether the hierarchy makes sense, and identifies potential areas where additional data may be necessary to confidently come to conclusions, then passes it to Steve
• (Removed) Steve performs final intensive review, including running a variety of formulae for data consistency and evaluating results against archival results for consistency of scaling
• (Removed) The Writer & Technician double-checks the finals. Any disagreement with what Steve passed is voiced and revisited. Once both are in agreement that it makes sense (or the suspect data is removed until validated), the data is sent to publication
• (Removed) The video editor, who is often technical enough to identify obvious oversights, performs a ‘passive’ QC while editing and flags any issues. If any are found, they go through Steve for review, analysis, and either sign-off as accurate or correction
• (New) Steve performs final quick review, including ad-hoc comparisons between CPUs from which we have a known relative % scaling, to rapidly vet additional information. Data with low confidence is removed rather than investigated. Not every single data point is inspected, unlike reviews

This allows the team to continue work on important next reviews without forcing us to skip more important upcoming projects, but still allows us to get helpful data out. You’ll notice that this approach cuts at least 2 potential internal reviewers from the process, including reducing the amount of times the SME looks over the data, and reduces the review of every single data point down to ad-hoc glances to see if anything “jumps out” as obviously bad.

If you see anything that looks out of order, you are welcome to email us at team at gamersnexus dot net.

We are hopeful that this will allow us to publish more data to help people make upgrade decisions, with a middle-ground understanding going into it as to the limitations of the process.

With all the helpful information on how to use these charts and the disclosures out of the way, let’s continue to the data set.

Long-Term Support CPU Charts (1H24)

LTS DATA SET: Zen 5 Review Cycle (pre-Arrow Lake)

The below CPU charts are those found from our long-term support list. These will be updated only a few times a year, but contain the most data for some of the older CPUs.

Production Benchmarks

This section contains the “production” benchmarks for workstation, creation, and development applications.

Blender 3D Rendering on the CPU

The above chart ranks CPU 3D rendering performance in Blender by best to worst (faster, or lower time, is better). This should help you identify the best CPUs for rendering in 2024; however, GPUs are heavily relied upon and would be a separate benchmark.

7-Zip File Compression & Decompression Benchmarks

The above charts contain our tests for 7-Zip file compression and decompression. If you frequently work with compressing and decompressing data, such as for large file transfers in compressed formats or for certain game loading tasks, these will give an idea for performance.

Adobe Premiere Video Editing & Rendering CPU Benchmarks

This chart uses the Puget Suite to benchmark Adobe Premiere for video editing and rendering tasks, focusing on CPU performance. Adobe Premiere cares both about the CPU and the GPU, and relies upon a strong combination (rather than biasing toward one component, like 3D rendering might do) for reduced scrubbing or playback ‘lag’ and improved rendering and encode performance.

Adobe Photoshop CPU Benchmarks & Comparison

This chart is for Adobe Photoshop and compares some of the best CPUs currently out for Photoshop. Testing is done with the Puget Suite and includes warps, transforms, scales, filters, and file manipulation.

Chromium Code Compile CPU Benchmarks

This chart looks at a Chromium code compile. It’s a CPU benchmark for programmers and developers, with the caveat that (like any of these tests), we can’t realistically represent all compile scenarios. Because we try to represent a wide range of possible use cases, we don’t cater too much to any one specialty. This should give you an idea for at least how this specific compile behaves. If your workloads resemble this, you may be able to assume some level of linearity.

Gaming CPU Benchmarks & Best CPUs

Even if your specific game isn’t represented here, the best way to use our charts for determining CPU differences is to look at the relative ranking across a number of games. Our intent with this approach is that you can determine the best-value upgrade (our reviews are very value-oriented) by comparing the typical or average gap between two parts.

In most scenarios, the CPUs will scale similarly from game-to-game, barring any unique behaviors of a particular game. If CPU A is better value in Game A, B, and C, it is very likely also better value in Game D (though not always). 

For individual per-game benchmarks, we’d recommend our Game Benchmarks section that deep-dives on new releases whenever we get a chance.

Dragon’s Dogma 2 CPU Benchmarks

This is for Dragon’s Dogma 2, which is one of the newest games in our CPU test suite. Dragon’s Dogma 2 is still getting regular updates, so the CPUs shown on this chart were all run on the same game version. The game remains CPU-heavy in the cities with dense NPC populations, which is where we test. We published a separate deep-dive benchmark of Dragon's Dogma 2 here.

Stellaris Simulation Time CPU Comparison

This chart evaluates simulation time in Stellaris using a late game save file. The number is represented in time, with lower being better. Faster simulation (shorter time required) is noticeable in 4X or Grand Campaign / Grand Strategy games where AI turn processing requires significant CPU effort. 

Baldur’s Gate 3 CPU Benchmarks

Baldur’s Gate 3 is tested at 1080p/Medium in the above chart, which allows us to see CPU scaling all the way up to the top of the stack for the best gaming CPUs.

We test in Act III in the city itself, near a market with a lot of NPCs.

F1 2024 1080p & 1440p CPU Benchmarks

This chart gallery is for F1 24 and includes both 1080p and 1440p results. Typically, in scenarios which remain primarily CPU-bound, we see roughly the same hierarchical lineup across both resolutions. The top of the chart truncates maximum FPS, which means that CPUs which occasionally bounce off of the GPU bottleneck will be less reliable to differentiate from one another (as they are externally bound).

FFXIV: Dawntrail CPU Benchmarks (1080p & 1440p)

This set of 1080p & 1440p charts is for Final Fantasy XIV: Dawntrail, tested at Maximum quality settings.

Rainbow Six Siege CPU Benchmarks

Rainbow Six Siege is a problematic benchmark as it updates frequently, causing entire wipes of our dataset. This is a list of results that were all run on the same game version. Unfortunately, we don’t have as much data for it as a result of the regular wipeout.

Starfield CPU Comparison

This chart is for Bethesda’s Starfield at 1080p/Low. Despite the game’s memeable launch, it remains a useful benchmark for CPU performance comparisons.

Total War: Warhammer III CPU Benchmarks

This gives us a look at a Grand Strategy / Grand Campaign Total War game, which tend to be CPU heavy. We’re using a battle for the benchmark.

Active CPU Benchmark Charts

ACTIVE DATA SET: AMD R9 9900X CPU Review

The below list of charts is our most heavily-vetted, most recently-published data set. Due to maintenance overhead and our focus on new content, we won’t updated it after every single review, but we will update this upload after major review cycles are fully complete. For example, if both AMD and Intel are launching CPUs across a 2-3 month spread, we’ll update this list at the end when we can breathe again.

These will lack some of the data found on the LTS charts; however, they may contain more recent data (such as newer CPUs).

There may be discrepancies between the LTS and Active charts. They are not necessarily comparable, and often are not. This is for reasons such as Windows version differences, game updates, and test platform changes.

Active Charts: Production

Rather than individually break them out into headings like above, we’ll just dump all the production charts below:

Active Charts: Gaming Benchmarks

And the same for gaming. You can find each game at the top of the chart:

CPUs Present on These Charts & Their Reviews

This table includes a simplified list of all CPUs on these charts, including links to the original GN reviews where present. Be advised that CPUs often have several follow-up pieces of content in rapid succession as the landscape changes. To keep things simple, we’ll just link the original reviews. You can still find the follow-ups on the channel.

Update Log

The below is an update log of changes to this page. The format is MM/DD/YYYY:

• 10/13/2024: Created page with initial dataset following Zen 5 reviews and preceding Arrow Lake reviews

Update process / house cleaning (externally visible, but used internally):

• Apply changes
• Append to update log
• Modify "Data Set" for both LTS and Active charts to identify the replacement dataset
• Modify current recommendations at the top of the page
• Append table of tested CPUs
• Update to link the latest Best CPUs article

This contains a recap of our recent face-to-face discussion with ASUS' Head of Customer Service

The Highlights

• This video and article follow our series of investigations into ASUS' RMA practices
• We aimed to get on-record, firm commitments of improvement
• We met with the Head of Customer Service, direct report to the CEO, and several team members

Table of Contents

• AutoTOC

Intro

We have answers from ASUS. Our latest video brings an in-person meeting with the company after publicly exposing flawed and potentially illegal handling of customer warranties and this article aims to rapidly recap the commitments that ASUS made during that discussion. The video itself is over an hour long and the (2) discussions spanned around or over 3 hours of total time together, so this list is just to get the fastest possible form of the commitments out there. The purpose is to inform the community in an efficient way, but also help keep ASUS accountable for its promises by condensing them into an easily referenced format long-term. This also helps us, and others in the community, remember what we exited the series with.

After our ongoing ASUS RMA investigation series, starting with the “ASUS Scammed Us” video, continuing with “ASUS Says We’re ‘Confused,’” and temporarily concluding with “ASUS Already On Government’s Radar for Warranty Issues,” we now have a for-now conclusion. We have asked ASUS to follow-up with us regularly as it continues to execute on its newly formed plan to overhaul its customer support and warranty systems. Likewise, we already have devices in their RMA centers under pseudonyms and plan to continue sampling them over the next 6-12 months so we can ensure these are permanent improvements.

The video is embedded below, followed by:

1. A list of the changes ASUS has committed to
2. How you can make use of those
3. A template for you to send to ASUS’ new email address established for anyone who wants to reprocess their prior warranty claims under the new conditions

Credits

Writing, Editing

Steve Burke

Video & Editing

Mike Gaglione
Vitalii Makhnovets
Andrew Coleman

ASUS' Committed Changes

Compiled By GamersNexus, resulting from GamersNexus coverage

(If you are another outlet reporting on our report, please credit our list and work. This list was not provided by ASUS. We created it.)

To recap several of ASUS’ firm changes:

1. ASUS now has a new inbox called “executivecare@asus.com” that they have created specifically to re-process prior RMAs that customers feel were unfairly classified, were misclassified, or charged for a service that should be free
2. ASUS has provided a template to copy and paste into your email to this address. We are showing it on the screen. You can visit gamersnexus.net to find a copy of this to copy and paste. We do not place third-party ads on our site. The link is below for the template.
3. ASUS has published a timeline for improvements: June 14th, today, is the publication of this email and template. ASUS has promised us an email this month with other changes.
4. ASUS has committed to refunds of service charges for unnecessary repairs which customers felt compelled to accept in order to have a warranted repair covered, such as unrelated or misclassified CID
5. ASUS has committed to refunding shipping charges in scenarios where a warranted repair was part of the RMA. For clarity, if a customer has both an out-of-warranty repair and an in-warranty repair in the same claim, shipping will be covered by ASUS
6. ASUS has committed to refunding labor and taxes related to these aforementioned qualifying disputes
7. ASUS has created a Task Force team to retroactively go back through a long history of customer surveys that were negative to try and fix the issues
8. ASUS has removed the power from the repair centers to claim CID. Now, CID claims must go through ASUS’ team. This will remove some of the financial incentive to fail devices. There still is one, but now it won’t be motivated as much by speed
9. ASUS is creating a new support center in the US. This will enable customers to choose between a repair of their board or a faster swap with a refurbished board. This solves an issue where refurbs were the only option in some scenarios previously
10. After over a year of refusing to acknowledge the microSD card reader failures on the ROG Ally, ASUS will be posting a formal statement next week about the defect, resulting from this series
11. ASUS will publish a more transparent repair report template in September of 2024
12. ASUS is changing the Advance RMA language to reduce emphasis on physical damage

ASUS Template for Re-Processing Old Claims

If you had a prior claim that previously, or now under the new changes ASUS has made, you feel was unfairly rejected, charged, misclassified (e.g. as Customer-Induced Damage, or CID), or otherwise potentially fraudulent under the FTC guidelines and Magnuson-Moss Warranty Act, you can use the below template to have ASUS fast-track you for reconsideration under the new terms and guidelines. Any legal claims for fraud should be handled separately through the correct legal entities; however, for those who just want the issue dealt with swiftly, going through ASUS theoretically is faster. If that is not the case, please alert us immediately with evidence at tips at gamersnexus dot net.

Template & ASUS Letter

"From the ASUS Service team, we deeply apologize for the issues with our service.

Our RMA communications and repair services have not consistently met the high standards we aim for. We want to make things right. If you’ve had a negative experience with our customer support, please let us know. We want to do better. Contact us at executivecare@asus.com and share your experience using the template below. We are here to listen and improve.

Your Name (as listed in your RMA): 
RMA Number:
Serial Number:
RMA application country:
Please describe your previous RMA dispute:
Supporting Documents (e.g., charged invoice, quotation notification, photos):
Additional Feedback (optional):

After submitting your claim, please allow us one week to review your case and respond. Our local service team will guide you through the appropriate next steps. We’re committed to transparency, fairness, and doing what’s right for our customers.  
We’re very sorry to anyone who has had a negative experience with our service team. We appreciate your feedback and giving us a chance to make amends."

The above is directly from ASUS. The italic text is the template. Paste that into your email to executivecare at ASUS dot com in order to get rapidly reprocessed.

MicroSD Card Issue & Misc.

ASUS will be making a statement next week about its ongoing microSD Card slot failures on the ASUS ROG Ally (which affected our original review unit). During the course of our discussion, we brought the failures up in a way that ASUS briefly, or maybe accidentally, acknowledged. This may be why the company is now moving to publicly acknowledge it. We think this is a good thing, as customers should know that it is a confirmed issue.

For the GamersNexus Warranty Toolkit for this or other warranty claims where you are uncertain what to do, please read here.

In the interest of keeping this post short and simple, we'll stop here. The video will contain more discussion. We will keep an eye on ASUS and continue monitoring the company. We also have devices out to other motherboard manufacturers.

If you have an issue with the ASUS RMA or Warranty process, please email us with evidence and a brief description of the problem at tips at gamersnexus dot net.

Thank you,

Steve Burke
Editor-in-Chief

We visited V-Color in Taiwan to get an inside look into its secret RAM-manufacturing processes

The Highlights

• V-Color employs a lot of custom-built automation to bin high-end memory
• As high tech as testing and manufacturing RAM can be, PCB separation can be a mostly simple process
• One V-Color robotic arm can test about 23,000 RAM modules per day

Table of Contents

• AutoTOC

Intro

Touring V-Color’s factory, we saw robots making thousands of sticks of RAM, numbering in the petabytes, and learned the secrets of RAM manufacturing.

The company’s Taiwan-based factory pumps-out memory using custom-built automation to bin the best overclocking chips for high-end memory. 

In this special addition to our factory tour series, we’ll see how SMT lines manufacture memory sticks for multiple large memory brands. We’ll see hundreds of testing stations scattered around, working around the clock to stability test RAM.

Credits

Host, Writing, Editing

Steve Burke

Camera, Video Editing

Mike Gaglione

Video Editing

Vitalii Makhnovets

Web Editing

Jimmy Thang

Automated Binning of Memory

We’ll start our tour of V-Color in its Automated Overclock Binning Lab. There’s no buzzword usage of “AI” here: These are just straight-up robots and automation, the old-school way -- except with ultra-advanced frequency and stability testing of modules.

Meet the backbone of this operation: These robots come in yellow and red variations, with each serving a special purpose. Naturally, the red robots are faster -- just like how red LEDs make your PC faster -- and both types set to work quality checking memory as soon as it enters the factory from the supplier.

The factory itself is extremely low-key: It’s located in Xindian, Taiwan, and is hidden away upstairs in this older, mixed-use office and residential building. 

The hallways leading to the advanced, ultra-clean, bright facilities we will see in this tour are deceptively dark, to the extent that we thought we were in the wrong place when we first arrived.

But just upstairs is this advanced OC lab.

This lab features high-speed binning robots and automation to determine the chip quality of each individual module that V-Color buys from suppliers like SK Hynix, Samsung, or Micron. They can process DDR4 or DDR5 memory and perform simple pass/fail analysis at a given frequency. Those that pass progress to burn-in testing on a per-module basis for stability, while those that fail are sent back to other robots for testing at lower frequencies.

Only two operators are required to keep this entire binning facility running, with the automation able to run 24/7. 

The screens outside of this lab show the different bins that V-color was testing at the time we visited. The large number at the bottom is the yield, or the pass rate, at that spec. A few of these are 100%, with one at 57.52%. That doesn’t necessarily mean the chips are defective, just that they can’t hit these settings. That particular test is configured at 1.18V and using timings of 17-19-19-39.

Other frequencies at the time we visited included 6400, 6200, 4400, and 7000+, but we’ll look at that later. 

Before we move on to burn-in testing, let's talk about how these robots actually do the binning.

The test is simple: The robot moves into position over the loading tray of fresh memory from the supplier, lowers, and uses 8 suction cups to pick up the individual chips. 

It then moves them into a prep board for a couple of seconds, then into the testing platform. The actual test typically takes about 15-20 seconds, with more time required for physical movement of the arm. 

The robotic arm then picks the chips back up and moves them over to either a pass or fail tray. Passes progress to the next step, with failures going back for binning at a lower spec or higher voltage. This particular test, including the time it takes the robot to move the arm between trays, took about 50 seconds per 8 chips. At 24/7 operation and assuming zero interruptions, that would mean this machine could complete over 1700 tests per day, which would be almost 14,000 modules per day for this one machine. 

The red machines operate much faster and have more complex movements so that the arm can maneuver around the chamber quicker. Its main advantage is that it has more testing trays, so instead of waiting for the test to complete before picking up the chips, it can alternate between two groups of 8 units. 

If you look closely when it’s placing the units down in the pretest tray, you’ll see two locations for those chips to rest. It also has two locations for active testing. This particular test was completing about 16 modules per minute, so at that pace, it could do about 23,000 per day without interruption. 

The yellow machines were the prototypes and can run for 1 hour at a time completely unmanned, at which point an operator has to load and check them. The new red machines can operate for 24 hours after loading, so this factory may be able to go even slimmer on staff for the night shift.

We asked a few questions while watching these machines execute their work with ruthless efficiency: First, we learned that they can test up to 4 timings, alongside voltage and frequency tests, to expedite the process. We asked what the error rate is of the machine itself, and we were told “0%.” It never makes an error. That’s impressive.

Other machines in this lab include a few things we couldn’t show and then an oven. This is also useful for NAND testing, as V-color can manufacture RAM and SSDs. The oven tests products at 98 degrees Celsius for endurance and stability.

After all of this automated testing, the memory takes a quick trip down the hall to the SMT machines -- or surface-mount technology line.

Assembling RAM Sticks: SMT Lines

This is the V-Color SMT room, where it operates 2 full-length lines for manufacturing memory. These lines consist of solder paste machines, pick-and-place machines, and machines that cut the memory PCBs. They can fully manufacture their RAM in these lines, and we even get a digital view of the parts progressing through the line.

Because they’re making RAM, it’s small and efficient and can largely be run with a single operator in the room.

Surface Mount Technology lines are used for almost every component at some point in the production process. And they never stop being cool.

SMT lines in some factories, like the MSI one shown above and below, also integrate wave soldering lines that help to attach through-hole components to the board. RAM at V-Color's facility doesn't require this step, but it's one of the ways SMT lines are expansible.

They also have fast-paced robotic arms that place components with deadly precision.

And the only human intervention is to keep a constant feed of reels upon reels of components, trays of GPUs or memory modules or silicon, and bars of solder.

Some lines, like MSI’s above, have extra steps for photography of motherboards or GPUs to check against customer RMAs in the future. We’ve also shown these at Gigabyte’s factory in Taiwan previously.

The longest SMT line we’ve ever seen is about 100 meters long and it was fully automated. Generally speaking, SMT lines will mount the smallest parts first and the largest ones later, which allows the machines to keep the speed high as they don’t need to dance around taller components (like inductors). Heatsinks are often placed by hand on motherboards.

V-Color’s line is smaller than the VGA lines by comparison, but it doesn’t need to be the same size. RAM is a relatively simple product: You have a PCB and a bunch of small caps and resistors, then the memory chips themselves. There’s not much to it.

The first step of the SMT process is to load the solder mask and the blank memory PCBs. The RAM goes through a machine that applies a thick paste to the board. This prepares the stick to get baked later, which solidifies the components to the PCB.

Next, a conveyor belt takes the stick away. 

Checking the monitor, we can see where each batch of memory is in the process. The blue blocks represent where the batches are. They move from right to left. The bottom right gives away the speed: This belt is moving at 82 centimeters per minute. We normally aren’t allowed to shoot these screens, so this is one of the shots we were the most excited about because it shows a commander’s view of the line.

But there’s another secret we get to show in this tour: A second screen. This one tells us the time break-down, and also outs how long we’ve held onto this footage. You can see that the assembly time is 38.68 seconds per boards, which is impressively fast. Currently, the line is processing 3,529 chips per hour and 70 boards per hour, averaging about 50 components per board. 

Remember, that’s everything: Each part on these reels and each part getting placed by these robots is counted here, not just the memory. Complex motherboards can be thousands of parts.

After going through the solder paste application, the RAM progresses to a machine that performs automatic optical inspection and begins the pick-and-place process. There are multiple stages of pick-and-place in this pipeline: Each pick-and-place machine pulls from reels of thousands of small components. 

In this shot, we can see the memory chips themselves getting placed on the PCB by the robotic arm, before and after each movement it’s optically scanning to ensure accurate placement prior to baking.

We also have a few shots where they paused the line for us to see the interior of the machines, but they had to be turned off for these shots. You can see each of the reels feeds into this like ammo belts, with the moving head actuated by gas lines and compressed air.

Ovens operate at multiple temperatures in this line, each baking the components securely onto the PCB. The hottest oven is the first one, meant to ply the solder paste for use, running at 260 degrees Celsius. 

The last oven is the coolest of them as the more sensitive components have been placed by this point.

At the end, the memory is spat-out 5 complete sticks at a time, consisting of all components and ready for separation. Compared to all these precision robots, PCB separation is a relatively simple process mostly involving a CNC router of some kind.

You can see as this technician loads the custom-made holsters for the 5-packs of memory, he gets the honors of pressing the giant green button. 

Instantly, just as fast as the button was pressed, the memory is sucked into the machine and you can hear the saw precisely separating the components and cutting the excess PCB off.

You can see the tiny blade vibrating as it cuts through the material, after which it spits the tray back out and the technician removes the excess circuit-board material.

All of the PCB powder is vacuumed out during the process and contained in a sealed container to be disposed of. Just these two lines alone can produce 10,000 sticks of memory if operating with the simplest modules and at maximum capacity.

Thus far, each stage of the operation has been impressively efficient, with only a few people needed in environments that, compared to some of the steel factories and chemical factories we’ve been in, are on the nicer side. But there’s one more stage to the process of memory manufacturing, and that’s burn-in.

RAM Burn-in Testing

This is the burn-in lab. And here, V-color tests every single stick of memory that it makes. This one room contains over 200 available test beds for memory. 

At a surface level, these are just test benches: They consist of a normal motherboard, CPU, and cooler. But on closer inspection, the memory is socketed into special risers, functioning as interposers that are used for validation and burn-in. 

Every stick that’s 8GB and under goes through a 1-hour burn-in test, with each stick exceeding 8GB requiring 1.5 hours for burn-in. The screens light up green for all the good modules and will flash errors for anything that fails the test. For the most part, bad chips should have been weeded-out in the first stage of the binning process, but it’s possible that entire DIMM sticks fail due to manufacturing issues in the SMT process. This station will help identify those.

In the event of a failure, the bad stick is taken for diagnostics by a technician. There are various ways to try and recover a bad stick: One is for desoldering and resoldering of any components that might be misaligned, although that would normally be caught by AOI in the SMT line. 

Another is with reflow ovens, where they can send the stick back through a local oven line just to reseat and reflow the solder. A final option might be determining that the stick is unsalvageable, such as with a bad PCB, and so V-Color will have technicians remove the valuable DRAM for re-use on a new stick.

RAM Overclocking Testing

The last station has the coolest job in the building: overclocking.

The company showed us its specialized one-off testing room for extreme OC memory binning -- and something caught our eye...

This station mixes manual and automated processes, where technicians also possess memory overclocking and tuning abilities. Ultra high-bin SKUs go through this line, contributing to some of the cost, and might be distributed to extreme overclockers or sold as a special line. 

You’ll notice that they keep memory cooling mounted to these stations: V-Color says that its A-die units perform best at 60C and below, with M die at its best at 80C and below.

As the memory modules exit these test rigs, they’re sent to automatic packaging and shipped to retailers and OEM customers. Memory is one of the most space-efficient manufacturing processes in the entire industry.

All of this is happening secluded upstairs in the mixed-use building, with the SMT lines literal steps away from the automated binning and the burn-in testing. That means any failures or reworks can be taken care of on the same day that they’re found, keeping process efficiency and throughput high.

And that’s the start-to-finish of how memory is made. The biggest unshown step is the actual silicon fabrication supply-side, but maybe that’ll be a future tour.

Be sure to check out the rest of our educational factory tour series as a video playlist below, and please consider supporting our tours (which are fully self-funded -- the factories are not permitted to provide travel assistance or pay for coverage) by buying something on our GN store.

Take a look at how InWin manufactures, assembles, and ships its cases and servers

The Highlights

• We take a look at InWin’s oldest factory along with its newer clean server and system factory
• Factory workers are getting better access to modern safeties, but still contend with hotter and louder conditions
• Automation is increasingly becoming a bigger part of InWin’s production line, but physical labor is still a large component of the company’s workflow

There are two very different approaches to manufacturing computer parts and servers in Asia’s sprawling mega factories: There’s the neurotically clean and obsessively sterile style of assembly and then the more traditional approach to manufacturing. In this article, we’re going to be looking at InWin’s case & server factories. Regardless of where you buy your chassis, the cases you purchase are made in places just like InWin’s. They’re grungy, dirty, and oily, and built for efficiency. We’ll be looking at InWin’s oldest factory and its newest in this tour.

The company’s cold-blooded efficiency gained from handmade customizations is what earned InWin the industry position and the money to make its clean, new, and nearby facility that we’ll take a look at first. The company’s clean server and system factory has hundreds of people work to bring the product from metal stamping down the street to its server assembly stations.

Credits

Host, Writing

Steve Burke

Video Production, Camera

Vitalii Makhnovets

Camera

Mike Gaglione

Writing, Web Editing

Jimmy Thang

Server Assembly Line

Pictured below is a server, delivered freshly assembled on an anti-static platter. These servers are built in a brand new factory for big clients we can’t name, but your web traffic has probably ended up on many of them.

Each server is not only built here, but tested too. And the testing process sounds like a jet engine.

We started our InWin factory tour in the assembly line which, just by price of CPUs and RAM alone, is probably one of the most expensive areas per square foot in the city of Taoyuan, Taiwan. 

Technicians take trays filled with terabytes of RAM and multiple AMD Epyc or Intel Xeon CPUs over to the conveyored workstation. 

From here, the technicians check the build order on the screen and start assembling the actual components of the server by hand. The cases are made down the way by an InWin factory that we’ll talk about later -- and actually, we’ve covered in the past.

There are 6 complete lines per floor at any given time for assembly, with each featuring its own SOP for the customer orders. Per day, this factory can turn through about 80 servers per line, maxing out at about 480 servers for this floor alone. 

A different line is configured for 2U systems due to changes in process. For full configurations with maximized RAM, an assembly line crew might socket upwards of 1,000 sticks per day. We’re betting the job application asks how strong your thumbs are.

As the systems are built, they are automatically carried down the conveyor belt and moved station-to-station. Eventually, the server is brought into a server elevator and brought out of sight.

But at GN, we’re still developing object permanence. So we asked the InWin crew to prove the server still existed once it left our sight. They told us we’d have to figure out a way -- and sometimes, that means we break the rules to get a shot. 

Standing atop a pallet stacker, Steve held the position above for about 5 minutes to get the shot below as we waited for InWin to send a server across. 

Until we realized something was lost in translation, and all they sent across was an empty turntable that a server would normally be on. Well, good enough.

Server Testing Facility

After the machines progress through this conveyor maze assembly line, they make their way over to this ginormous, high-power testing room. 

Up to 800 1U servers get burned in just in this room alone. As for its power capabilities, the room is equipped with 80 total test lines, each  capable of delivering 5,000W. The total power handling capacity of the testing floor for the factory is a staggering 400kW.

Total burn-in time is 2 hours per server at 10 units per rack. And one guy controls almost all of it -- meet the man with the most powerful index finger in the world, trained from entering BIOS hundreds of times per day.

The test racks in this area cycle through I/O testing to make sure all ports fire, CPU burn-in, cooler burn-in, and even testing the entire RPM range for the fan. And it sounds kind of like a deafening aircraft carrier when they all fire at once.

And no, OSHA doesn’t exist here. That’s also why Steve was allowed to climb the pallet stacker.

The testing suite is largely developed by the factory itself, with software designed to alert technicians to bad components or test failures. Anything that fails goes to the repair station, where technicians will identify the bad part and rescue the rest of the server. Some bad components, like sticks of RAM, can be sent in for board-level repair by the original supplier -- especially since many of these factories are local to Taiwan, making it logistically easy.

Other testing is done in a side room for low-volume custom builds. 

Steve was briefly assigned a job he couldn’t possibly screw up: Move the monitors up and down with a button. These stations are also used for documentation and photography to keep a record of serial numbers and parts.

The servers are loaded onto pallets when testing is complete, then prepped to move to packaging.

Water Cooler Testing

There’s one more secret room before we go to packaging, and it’s pictured above. 

This brand new facility is for water cooling testing on servers. The huge amounts of perfectly installed plumbing in this room allow for up to 38U or 20KW of servers to be tested at once. 

The water pumps through these huge hoses, with pressure gauges placed throughout to help check for leaks. They’re also checking the water temperature here. This can be configured to test water cooling for entire server racks -- or multiple -- for complex systems. 

Or they can test a single unit at a special station, pictured below:

All the water is pumped up to a giant radiator and set of tanks on the roof, which we were able to capture with our drone shots:

Packaging Robots

After this entire process, everything goes to packaging -- which is a whole beast on its own.

The images above showcase the packing station. Here, pallets upon pallets of servers are unloaded onto conveyor belts. 

A technician screws all the panels together, then performs final quality checks and inspection for all components. 

A team manually packs the servers into foam and boxes.

Despite being maybe one of the least sophisticated parts of this factory, the automation around a roll of tape is surprisingly satisfying: Each box gets auto-taped, rotated, and taped again, and although we couldn’t name any of the clients earlier, there might be hints in the image above as to who some of them are. But we’re not sure...

Continue reading below to see the grungier painting factory.

InWin Case Factory & Painting

A quick car ride away, just down the street, is InWin’s longest-standing, traditional factory, which underscores some truth about factory life.

These floors are kept dark to help combat the heat and humidity of Taiwan combined with the heat of ovens baking cases at nearly 200 degrees Celsius. 

In this old factory, cases are automatically moved along a ceiling-mounted conveyor line using hooks. One by one, the cases climb to get painted and then baked. 

The walkways are crowded with machines and massive equipment -- the claustrophobic wouldn’t do well here. There’s also the everpresent sound of paint waterfalls surrounding this area.

On the scale of factories we’ve been to, the presence of at least respirators is better than we’ve seen in some others -- but it’s certainly not conditions we’re used to in the modern West. Between the noise and the air of the many non-robotic painting factories we’ve been to, floor managers at various companies have told us that it’s been hard to keep staff around for very long. 

Robots have been slowly filling-in the gaps for this former manual workforce, as we showed in our previous automated painting factory tour. Maybe in instances like these, robots are for the best.

Workers prep the cases and paint them as they’re carried from one paint booth to another. The hooks are mounted in a way to minimize obstructions to the paint, workers manually mask some parts of cases, and they’re brought at a fixed pace through the painting process. Later in the process, more automation is introduced for the generalized coatings, whereas the earlier manual processes focus on hard-to-hit spots. 

These waterfalls help wash the overspray down into a catch basin mixed with paint and water, amounting to millions upon millions of liters of use annually. The factory is able to recapture and reclaim much of the water and squeeze the paint out -- which is something we showed in a prior Lian Li painting factory tour -- which reduces usage. 

Pipes are exposed outside of the building that allow the Taiwanese government to perform unannounced audits of the filtration quality.

The entire painting process for each case takes about 40 minutes, and the baking process starts first at 75C to slowly harden the shell, then ramps to 180 degrees, and finally 200 degrees for powder coating. 

InWin is slowly modernizing all of its factories to be more like its new server facility, and many other factories in Asia have been moving increasingly to robotics -- something we’ve been showing more and more over the years. And although the grit of a machine shop and tooling factory will be blended with the sterile precision of robots, the frenetic pace of these factories will never go away. 

In our time at these factories, we saw constant movement and an unspoken understanding between everyone working who needed to be where and when. 

To see the differences between the quiet and collaborative design rooms filled with inspirational artwork, bright colors, and art books packed with future case designs and the factories in which they’re made is a privilege of ours, and we’re excited to continue reporting on what’s really behind the products you buy in the PC industry. If you want to learn more about the front-end of product design prior to manufacturing, check out our tour of InWin’s design room.

After all of these steps, the products finally get packaged outside and shipped in the millions to PC builders and data centers around the world. In our upcoming tours, we’ll be looking at RAM programming and automation, bearing manufacturing, and lasers -- so make sure to check back regularly.

How component manufacturers coat and color aluminum with tens of thousands of kilowatts of electricity

The Highlights

• Used to create a high quality protective coating on PC components
• The most expensive of the available methods of placing a coating on PC components
• Requires highly caustic chemicals and tens of thousands of kilowatts of electricity

Computer components come in hundreds of different shapes, sizes, and colors.  However, some components come out with a more complex coat of paint than others, like the aluminum Yeston CUTE PET case , while others use a simpler paint job. When component manufacturers decide how to coat their products, they typically choose from anodization, painting, electroplating, and electrophoretic deposition, with the cost listed from high-to-low. We’ve observed Noctua coolers getting coated black in our factory visits, using cathodic electrodeposition, and we’ve also seen the robotic arms responsible for painting case side panels like the Lian Li panel made for Digital Storm’s bolt . This article will look specifically at anodization at a third-party factory that Lian Li uses for coating its more expensive components.

The anodization factory we visited is located in northern Taiwan. The factory manager gave us a rundown of the process, from cleaning the parts, stripping them down with chemicals, electrifying them in a bath of acid, coloring them, and finally packing them back up to be returned to the manufacturer.

Editor's Note: This was originally published as a video in April 23 of 2020. This article conversion does not change any of the original editorial content, but is being dated with the article entry date to float it to the top of the website.

Credits

EDITORIAL, REPORTING

Steve Burke
Patrick Latham

video

keegan gallick
Andrew Coleman

WEB Editing

Jack Reitman

How Anodizing Works

Visually, the process of anodizing parts in a factory looks similar to electroplating. Anodizing is fundamentally different in that it doesn’t involve coating one metal in another. Instead, it’s a method of adding an additional oxide layer to the existing metal. Anodization works on many different metals, but is commonly associated with aluminum, which is what this factory specializes in. Aluminum forms an oxide layer on contact with air that naturally protects the metal and prevents further oxidation of the metal underneath, but this natural layer is extremely thin--anodization allows growing a much thicker protective layer that’s easier to dye.

Mounting the Components to Be Anodized

• Mounting on aluminum racks
• Mounting on titanium racks
• Components to be anodized

The first step of anodization is to fix the aluminum parts onto a conductive metal rack. Two different types of racks are used, depending on the parts that are being coated: titanium racks are stronger and can hold more parts, but the conductivity isn’t optimal, so these are used only for functional parts that are being anodized for protection rather than for aesthetics. Aluminum racks offer better conductivity but hold fewer pieces, and so are used for pieces that are designed to be seen, like Lian Li’s external panels of a case. 

Any point where the hooks contact the piece won’t be anodized, so the aluminum hooks are custom-made for each new type of part to minimize contact, while the titanium hooks have a more standardized design. The factory technicians figure out the best spot for each product to attach to the hooks, typically choosing unseen locations like inside of a screw hole. The hooks themselves are anodized during the process, so they won’t last forever, but they are reusable. Factory technicians bring empty titanium racks back from the packing area next door to start the cycle again.

Aluminum comes in many varieties, but the factory most commonly works with 5000 and 6000 series alloys. When asked if some are easier to work with than others, we were told that harder alloys tend to come out grayish, while softer ones produce a better color due to the change in conductivity.

Chemicals and Cleaning

• Part cleaning
• Degreaser vat
• Running water tank (pictured in back)

Once the parts are secured to a conductive rack, they next have to be cleaned. It’s critical to the operation that the surfaces are clean and free of any fingerprints or other contaminants for the rest of the process, so they’re first cleaned with an oil, and then dunked in a bath of degreaser for 35 minutes, followed by yet another bath of water to rinse off the degreasing fluid and the contaminants that it removed. These tanks are held in the middle of the first room, surrounded by a short concrete curb to hold in splashing and overflow from the water rinse tank, which is being filled with fresh water constantly. The insulated bumpers above these tanks are just to hold the metal poles from which the racks are suspended--electricity isn’t involved until later in the process.

Chemical preparation for the anodization process begins after initial cleaning. The 7 tanks for this stage are lined up against the very back wall of the factory, securely away from the entrance. The first step is pure water cleaning, then the first chemical step is dunking the parts in an alkaline mixture of sodium hydroxide, also known as lye or caustic soda, for about ten seconds. This mixture is kept somewhere between 60 degrees and 70 degrees Celsius and the tanks are visibly steamy, producing a vapor that must be pulled away by a fume hood placed specifically over this end of the line. Lye can cause serious chemical burns, so caution must be exercised here.

The lye bath is necessary to dissolve the natural layer of aluminum oxide, which lacks the pores required to let oxygen in and grow the oxide layer. Aluminum oxide is amphoteric, meaning it can react with both bases and acids. Sodium hydroxide reacts with it to produce water, aluminate, and heat. Sodium hydroxide also reacts with the raw aluminum underneath to produce aluminate and hydrogen gas, hence limiting the bath to ten seconds. Surface finish is determined at this stage--to make the end result shinier or duller, special chemicals can be used here.

After the ten-second dip, there are five separate water baths, one next to the other, to wash away any remaining traces of sodium hydroxide. Afterwards comes a mixture containing nitric acid. Lye does a great job of eating away aluminum oxide, but aluminum is commonly used as an alloy, and it may not be as effective on the alloyed metals. Diluted nitric acid is used to wash away the residue of other metals in a process known as “desmutting,” followed by yet another series of water baths. Before it’s diluted, the nitric acid is delivered in jugs at 68% concentration--any stronger and it would be considered potentially fuming nitric acid, possibly forming fumes on contact with air, which is undesirable and a safety hazard. The time between putting the parts in the lye bath and finishing the last of the water rinses is approximately ten minutes.

The parts have been completely stripped of any aluminum oxide coating at this point, so it’s vital to prevent contact with the air as much as possible. There are several “standby tanks” of water here for keeping aluminum parts submerged in case the production line gets backed up.

Blasting Electricity at Metal

• Anodization line
• Parts submerged in sulfuric acid solution with electric current passed through
• Copper anode

Finally, anodization can begin.  As with electroplating, the aluminum is submerged in fluid, in this case a solution of 20% sulfuric acid and 80% water held between 20-25C, with an electric current passed through it. The racks containing the aluminum parts are suspended on a metal frame wired up with massive copper strips, through which the current is passed. Wedge-shaped copper slugs at the ends form the contacts. A crane is operated manually at a control station nearby and moves back and forth over the tanks, lifting the frames in and out of the appropriate tanks. 

There are only four anodization tanks in this particular line - the rest are for rinsing and coloring. Unlike electroplating, the metal to be anodized is the anode of the circuit, hence the name. The electric current splits water molecules into hydrogen at the cathode and oxygen at the aluminum anode, forming visible bubbles and causing the aluminum to oxidize. The acidic electrolyte bath eats into the oxide layer without completely dissolving it, allowing oxygen to reach even deeper and oxidize even more of the metal, and forms a much thicker and more porous structure than would naturally occur. The end result is an inflexible oxide crust that’s not electrically conductive and is less thermally conductive than the aluminum underneath. 

As for power, this line of tanks was running at an impressive 25KW. We were told that the actual anodizing process length is 20-40 minutes, with darker colors taking longer. Black takes about 40 minutes, white or semi-transparent coats take about 20 minutes, and the rest is in between. This process can be started from either side of the line, depending on logistics each day. Parts are rinsed 6 times in water and then colored.

Coloring and Quality Control

• Dye tank
• Dye vats
• Waste water return

Anodization doesn’t directly color aluminum, but it does provide a nice open porous surface for dyes to stick to--darker colors require a thicker oxide layer, and therefore more time to be anodized. The main production line only handles these two colors, black and white, as they’re the most popular. Other colors or smaller batches of parts are done manually in smaller tanks lined up alongside the larger ones. This factory offers thirty different standard colors--custom colors are also possible, but require much more effort. The dye tanks are held at 110C-160C, but take only about 15 minutes to preheat in the morning. Parts must be submerged in them for 5-10 minutes to be dyed. After coloring, the parts are rinsed three more times in water.

Every step of the anodization process requires large volumes of water, which must be replenished throughout the day. Anodization, as compared to painting, produces relatively low waste since the aluminum-containing byproducts filtered out of the water, like aluminum hydroxide, can be used by other industries and resold. The water drains out of a downspout at the front of the factory, and surprise government inspections to check the water quality are commonplace. We were told that it is easy to meet these requirements because the chemicals used are not too toxic.

After coloring comes QC. Color is checked against a reference to make sure it’s within an acceptable range. The most common defects come from parts moving around or bumping into things and damaging the finish.  If the racks aren’t moved carefully, the parts can scratch each other. Smaller pieces are easier to deal with, while larger ones cause more problems. The factory can afford to be picky about blemishes because almost all defects can be fixed by re-stripping the oxide coating and simply repeating the process from square one, but if any serious damage has occurred, the parts are sent back. The customer sends in an exact number of parts for anodization and expects all of them to come back, so the factory must do its best to reclaim every defective part. 

The final step is packaging. The factory doesn’t sell these parts themselves, so that means securely wrapping up the parts in bulk and delivering them back to their customers, like Lian Li.

As mentioned above, this is just one of several processes case manufacturers use for surface coating, depending on the requirements. Anodization is the preferred process for externally visible aluminum parts, but for a smoother and less expensive finish they use electrophoretic coating, or for vibrant and varied colors they use paint. Each of those processes has their own interesting differences.

Our factory tour playlist (that includes related processes like painting, electroplating, and electrophoretic coating) contains over 20 factory tour videos in Taiwan and China alike.

We've tested dozens of CPU coolers. These GN Mega Charts include our air vs. liquid benchmarks & more.

The Highlights

• The data on this page is intended as a long-term support / long-term reference dataset
• We will update this page a few times a year. The URL is permanent for our full cooler bench results
• This page will not cover pressure tests, installation, or mounting. We will only cover data. Check the reviews for more nuanced opinions or concerns
• This is a part of our Mega Charts series. You can check our Features page for more of these
• Please consider supporting this effort

Table of Contents

• AutoTOC

This is another of our GN Mega Charts series! You can find all Mega Charts on the Features page, such as our CPU power consumption mega charts.

This page will be regularly updated with the latest of our CPU cooler benchmark performance numbers, including links to CPU cooler reviews, comparisons against stock coolers (and discussion over whether it’s worth upgrading something like an AMD Wraith or Prism cooler), and more. The page will run on a slight latency from our latest reviews, but will be updated a few times a year with larger charts than you’ll find in those latest reviews. There are a few reasons for the truncated reviews charts, but the main one is simply legibility: Particularly in video format, it’s hard to fit much more than ~20 items (with 2-3 bars each) while still maintaining usefulness, like on mobile. The other is to focus the discussion or viewer in the short time a chart is on screen. This page doesn’t have those limitations.

We’d like our audience to think of these Mega Charts round-ups as an “LTS” (or Long-Term Support) version of our data. They don’t get updated constantly, but allow us to maintain publication of quick-reference material and a stable set of data. This is a permanent page. This URL will always contain the Mega Charts round-ups for coolers.

Credits

Test Lead, Testing, Writing

Steve Burke

Testing

Mike Gaglione

As a note, we’re currently working on revising our CPU cooler test methodology to move to a new platform. When we do and eventually update this page, we’ll probably split the sections into “chaptered” headers for old and new test benches. For now, all of this was conducted on the same bench.

Best CPU Coolers | Benchmarks Chart Gallery

There are important notes below this, but we’re going to provide all the charts right up front to make this as convenient and accessible as possible! If you want to learn about the methodological choices, the limitations, and the individual coolers (with links to their reviews), you can continue down past the charts gallery. We’re not yet sure how we’re going to change this Mega Charts page when we refresh our cooler benches soon, but we may end up keeping both sets of data here (with the 2019-2023 bench lower down and the new bench higher up). 

In either event, you can find an update log at the bottom of the page. You can learn about our methodology here. Presenting the gallery first assumes some knowledge of the user. If you are unfamiliar with the nuances, we’ll provide some resources below it. Here it is!

Intel 14900KF CPU Cooler Benchmarks

CPU Thermals: Intel 250W Noise-Normalized & 100% Fan Speed

CPU Thermals: AMD 200W Noise-Normalized & 100% Fan Speed

CPU Thermals: AMD 123W Noise-Normalized & 100% Fan Speed

CPU Thermals: AMD 68W Noise-Normalized & 100% Fan Speed

VRM Thermals: AMD 200W Noise-Normalized

There are some very specific notes for this one that we’d encourage you to read before going too far with this information. Those are below.

Using This Page

This page will contain all of the updated information, as explained above, as we continually roll-out reviews. The reviews will always contain the newest and freshest data, along with the analysis, but this page will serve as long-term reference material. This page will not contain individual cooler analysis. This page also won’t talk about mounting quality, installation, pressure maps (which are a major part of our reviews), or individual gripes, criticisms, or praise of coolers. Instead, it will be fully data-focused (again, sans pressure). If you identify a cooler you’re interested in on this page, we’d encourage you to check our review. Remember that good thermal performance does not necessarily mean it’s a good product.

For coolers listed, you can Ctrl+F to find whether the one you’re interested in is included in any of these charts. We’ll have a table below that links to Amazon and Newegg pages for the coolers (with affiliate links) and links to the original reviews, if present.

Mega Charts pages should serve as bookmark-able reference material. The URL will remain the same, so save it for regular reference. An update log will be maintained at the bottom. As we are currently an ad-free website, we ask that if you find this page routinely useful and want to support our maintenance of it, that you please support our efforts. You can do this a few ways:

• Patreon: This is a monthly subscription of any amount to help fund our continued efforts
• Direct donation: This is a one-time, direct donation of any amount and will ensure we incur the lowest fees
• GN store purchase: This is the best way to support. You can get a PC building accessory or tool, a mouse mat, a t-shirt, glassware, or any other item you find useful while also supporting us

Testing Methodology

Results, as always, are fully dependent on methodology. One of the most often user errors we see is people saying “my CPU runs at 80 degrees,” then looking at a 60-degree result on our chart as a clear improvement. Keep in mind that unless everything is the same, they really can’t be compared this way. Thermals are one of the most isolated tests we run, and as opposed to a 3D Mark score, they cannot be compared cross-platform to yield any useful information. They can be compared intra-platform (as we’re doing here) to determine the most likely “best” CPU coolers, or to do tests like air vs. liquid coolers for PCs, and then you can use that information to apply to your build.

GN CPU Cooler Test Bench (2019/20-2023)

This is the bench we used for all of the testing shown here. The bench was built in ~2019 and fully publicly launched in 2020, following about 6 months of internal data validation.

Platform Choice & Heat Loads

The heat loads represented in this testing include:

• ~190-200W
• ~123W
• ~68W

This was conducted on an AM4 platform with an X570 ACE motherboard that has now been in service for about 4 years. We’ve been impressed with its resilience to non-stop thermal stress. Given its survival, we’ll be using an X670 ACE board for our new AM5 benches.

Measurement & Environment

Again, this is described in detail in the methods piece. The basics are below:

Thermals: Thermals are measured with a mix of HWiNFO64 Engineering Edition and a thermocouple reader. We use a calibrated K-type thermocouple to monitor ambient temperature at the inlet of the CPU cooler, then subtract that from the averaged Tdie readings (at steady state) to produce a “Delta T over ambient” number. That’s what’s presented in the charts. We find that, within the small deviations in ambient air temperature in our test environment, this is functionally completely linear. If we see a +/- 0.5 degree Celsius fluctuation in the course of the test, this approach helps to smooth that change to isolate only the true impact to thermals from the CPU and the cooler. Testing is done in a 21C environment that is carefully controlled.

Flatness: With our original methodology, flatness is measured point-to-point with a precision needle instrument that measures the distance in microns from a known 0-point (calibrated annually with certification). The instrument allows us to measure the point-to-point deviations in surface flatness. A “flatter” surface, generally, is not inherently better; however, our test mostly looks for major craters or pitting in the surface of the coldplate. The best example of where we found an issue was in the Corsair A500 CPU cooler, which had some large gaps that caused it to actually sheer paper when mounted on top of a sheet of it.

Power: Power is arguably the most important aspect of cooler testing, as it’s what dictates the end result. We measure power into the EPS12V rails constantly throughout the test with a clamp. This is primarily for internal use so that we can ensure the bench is not drifting with time and that BIOS hasn’t applied some sort of unexpected voltage somewhere. EPS12V cable measurements allow us relatively ‘true to life’ measurement of the current, just with VRM efficiency losses included. One thing we’ve noticed is that power does drift with higher temperature “coolers” (technically, the CPU runs hotter, but that’s because of the cooler): In such instances, this is part of the test. We’ve found that there is typically a 4% reduction in power leakage for every 10-degree reduction from the cooler, although this isn’t a perfect science. For instance, one of our tests has a 48-degree result at 191W. Adjusting the fan speed to worsen the performance to ~65 degrees (over ambient for both) led to a 206W measurement, which is approximately a power reduction of about 7.28% for a ~17-degree reduction in performance.

Acoustics: With this older methodology, we were still using our room acoustic measurements. These are being retired as we replace them with our hemi-anechoic acoustics chamber, which was a ~$250,000 build project that we filmed. You should check it out! This methodology uses a room measurement in a ~26-27dBA noise floor at 20” measurement. The bench is fully passive other than the cooler.

Limitations

This testing, like all testing, has limitations. There are aspects of cooler performance not represented here, and likewise, some results could be more specific to the bench and methods used than other results. We always want our audience to be aware of the limitations of testing, as there is no such thing as a fully representative, perfect test, so it’s all about knowing how to leverage the information within the limits of its capabilities and usefulness.

Although much of this is discussed in our methodology piece, a few basics are below:

First, this testing was done on an AM4 platform. Generally speaking, methodologically, what matters most is the amount of power going into the chip and the heat load being tested. In theory, two parts that both draw 200W should have a somewhat familiar or largely similar hierarchical ranking of coolers. But this isn’t always that clean: A few of the biggest limitations are IHS shape and design, die or chiplet location, IHS size, and mounting hardware included with the cooler (for instance, if a manufacturer includes a to-spec kit for AMD, but an undertorqued spec for Intel).

One of the reasons we’ve moved on from our AM4 standardized cooler test platform and are working on rolling-out an AM5 + LGA1700 platform is to do with IHS curvature and shape. AMD’s IHS is relatively flat, all things considered, and Intel has a slight curvature to it. Intel also uses a more oblong design (in the LGA1700 generation), whereas AMD is generally square. One of these isn’t necessarily better than the other, but they can have an impact on cooler performance. In some cases, it’s a dramatic impact. In many, particularly with lower-end coolers, it’s not as noticeable since less engineering goes into matching the coldplate to the IHS. It’ll be double the work to test both AMD and Intel in the future, but doing so will account for any coolers that tune for IHS design.

In other words, while you can (generally speaking) assume the same or similar hierarchical ranking of “best” CPU coolers from these charts, you should also be mindful of differences in your intended use. If you intend to build an LGA 1700 system, then the ranking may be altered. Barring a major oversight by the cooler manufacturer (which, of course, is half of our job to look for -- hence adding both CPU types), good performance on AM4 should lead to good performance on LGA 1700. But it’s not always the case, so that’s the limitation of testing here.

A lot of times though, we’re talking about single-degree differences. For many people, these don’t matter.

Let’s get into the charts from our galleries earlier, but with some extra information.

List of CPU Cooler Comparisons: Best Coolers for Gaming PCs

The below table contains all coolers tested and present on these charts.

CPU Thermals: 200W Noise-Normalized & 100% Fan Speed

The above charts gallery includes data from our ~200W CPU heat load, representative of many modern high-end gaming CPUs. These are the most power-intensive tests we ran with the AM4 methodology.

The charts include a noise-normalized test and a 100% fan speed test (irrespective of noise): in the noise-normalized test, we define whether the coolers are liquid or air next to their names; in the 100% speed tests, we define their noise levels at that speed. In all tests, the stock (included) fans are used unless otherwise noted. Some coolers include an extra fan and are noted if so. We average the RPM between all fans present for simplified representation.

Remember that 100% fan speed tests do not control for noise levels, and so generally speaking, with a similar radiator or tower, faster (louder) fans will “win” on 100% charts. The noise-normalized result is what we generally weigh as the most important for a cooler, as it shows us a more like-for-like scenario. 100% results are still useful though, as they give an understanding for the maximum performance potential. In those charts, the “efficiency” of the result matters: You can look at the noise level (next to the cooler) against the temperature. If there’s a result that scores better than another while being quieter or the same, that result is more “noise efficient” (by our terminology) for the temperature we get.

CPU Thermals: 123W Noise-Normalized & 100% Fan Speed

Important: We are now moving to lower heat loads. As heat load reduces, the differences between high-end coolers diminish. It is not fair to judge a high-end cooler’s peak performance based on an underwhelming heat load, sort of like how it wouldn’t be fair to judge a CPU based on a load that’s GPU-bound; however, we still include those results because they’re real-world numbers that may help put into perspective the possibility of over-buying your cooling solution. Typically, the biggest advantage from an “overbought” cooler is that you can reduce noise levels more notably.

CPU Thermals: 68W Noise-Normalized & 100% Fan Speed

Now for the lowest heat load charts. This data also gives us a look at stock CPU cooler performance vs. ‘aftermarket’ coolers, like the AMD Wraith Prism and AMD Wraith Spire stock coolers vs. the Thermalright Assassin Spirit, Cooler Master Hyper 212, and more.

VRM Thermals

Now we’ll get into VRM thermals. These results can be wildly variable depending on how the platform is tested. VRM thermals are influenced by coolers primarily from the position of the fans: If a liquid cooler has a VRM fan atop it, that’ll obviously benefit the VRM thermals. We showed that here. Likewise, if an air cooler uses 140mm fans that sit low in the heatsink, they’ll pull air over or push it into the VRM heatsinks to cool the MOSFETs that we measure.

Liquid coolers are side-mounted on the test bench in such a way that would represent a top-mounted orientation (intake) in a case. If you intend to front-mount your AIO, the fans will not benefit the VRM -- likely not at all -- and you’d need case fans to help it. Air coolers can be directly compared here, but liquid coolers can only be compared under the assumption of the same mount as tested.

Likewise, for noise-normalized testing, you’ll see some liquid coolers seriously struggle with VRM thermal performance. The EVGA CLC 360 is a great example of this in our charts, as its pressure wanes significantly after the reduction for acoustics, causing it to lose pressure through the radiator and lack power when it hits the VRM heatsink.

These charts are also most useful for low-end motherboards where VRM thermals are a concern. A lot of modern boards are overbuilt, which reduces the necessity for ancillary VRM cooling.

Update Log

The below marks our updates and change log for this page. The format is in MM/DD/YYYY:

• July 14, 2024: Added data for the D15 G2 on 200W AMD. Added data for the Arctic Freezer 36, Peerless Assassin 120 Mini, ID-Cooling A620, ID-Cooling A720, Cooler Master MA824, and Thermalright Frozen Prism 360
• July 7, 2024: Added Intel LGA1700 (14900K / 14900KF) 250W heat load, following the Noctua NH-D15 G2 review
• December 2, 2023: Initial creation of the page with first dataset following our Best Air Coolers round-up

CPU all-core power consumption benchmarks across generations

The Highlights

• This data has been collected from our CPU power consumption testing
• Data here is directly comparable and can be cross-compared between charts on this page
• CPUs have been split by release year and/or generation to make filtering easier. Ctrl+F for the CPU of choice
• This is a large, ongoing effort and will get updates at this URL permanently
• Please consider supporting this effort

Table of Contents

• AutoTOC

Intro

This is part of our long-term reference series of component charts. We will roll these out over time.

This page will be updated regularly with our latest power consumption figures for CPU testing. There will be a slight delay to incorporate the latest data following a CPU review, so the charts on this page may not be as up-to-date as our latest CPU review; however, it will contain more data than is presented in the reviews. If we were to make a software analogy, think of this page as the “LTS” (or long-term support) version of the charts: They won’t be updated as often, but will contain the most stable set of data. 

Credits

Test Lead, Testing, Writing

Steve Burke

Testing

Patrick Lathan
Mike Gaglione

CPU Power Consumption Charts Gallery

There are important notes below this; however, we want this page to be as useful as possible, so we're putting the charts right at the top in a gallery format for quick reference. Please scroll down and read the methodological details for further information, and also note that the Legacy & Miscellaneous chart has an important bit of information about BIOS controls and test parameters accompanying its section at the bottom of the page. There is also an update log at the bottom. Here's the quick-reference gallery!

CPU Power Efficiency Chart

The above is the current (as in 'today,' not as in P=VI) power efficiency chart. Unlike our above (and below) power consumption snapshot tests, this chart cannot be compared across multiple test generations. That's because it relies on a unit of work (a Blender render) to complete for the calculation, and that unit of work gets updated for our review cycle every 1-2 years. Everything on the chart is directly comparable, but it cannot be compared to older versions of itself. This gives us a measure for a known and fixed unit of work (the render) versus the power required to complete that unit of work. As an example: Something that's faster and draws more power could be less efficient than something that's slower, but ultra low power.

Using This Page

This page contains a mix of test methodology information, limitations of testing data, and the data itself. We also present links and resources for the original reviews (or related content) and to the Amazon or Newegg pages, which contain affiliate links to help support this page.

Of course, we'd recommend reading the limitations and methodology section first. Once done there, we'd recommend Ctrl+F to find the CPU you care about. that will jump you to a table naming the CPUs in each chart.

We intend to update this page quarterly. The results are dependent on methodology. As long as the method doesn’t change for each chart, it can be appended going forward.

We suspect these long-term reference pages will get higher traffic volume than typical. The intent of these pages is to serve as a quick reference, permanent URL for performance data. Because we currently are ad-free on this website, we ask that if you find this page regularly useful, you please support our maintenance of this information. You can support us a few ways:

• Patreon: This is a monthly subscription of any amount to help fund our continued efforts
• Direct donation: This is a one-time, direct donation of any amount and will ensure we incur the lowest fees
• GN store purchase: This is the best way to support. You can get a PC building accessory or tool, a mouse mat, a t-shirt, glassware, or any other item you find useful while also supporting us

Testing Methodology

Unless stated otherwise, all power tests on this page were conducted using the same testing methods. The test benches have changed (naturally) for each CPU tested; however, because we isolate to only the EPS12V cables and test at a full and fixed load, the only meaningful impact to power consumption would come through motherboard changes.

Methodologically, we control all motherboard power settings (unless otherwise stated for the DUT) to match the official guidance from AMD or Intel. That means boards which run MCE out-of-box will be tuned down to operate at the proper guidance and closer to the commonly understood “spec” (although that specific word is arguable) by the community. 

We are advancing and changing our testing processes for power all the time, but for purposes of maintaining a directly comparable dataset, we maintain these tests long-term.

Software Workload

The test uses a full, all-core Blender rendering workload using the Cycles renderer to load the CPU to 100%. We use an “impossible to render” GN logo frame with high sample count and resolution.

Timing & Turbo

This set of tests includes a 5-minute wait period (under load) before taking a steady state power figure. This does a few things: 

First, it ensures we’ve had some time for leakage to set in and for power to stabilize. Power will bounce more at the beginning of a workload, especially on CPUs that may have features like Precision Boost or may boost to a particular thermal ‘value’ for Tdie. 

Secondly, and most importantly, it ensures we are measuring after any Turbo time limits on processors that have them. We do this to make sure we are consistent in measurement. Different tests are performed for the Turbo power consumption (this occurs within the first 45-50 seconds of boost on older Intel CPUs). For all-core, steady state testing, we measure only after the CPU has stabilized. The Turbo/Tau tests are not on this page yet, but again, that primarily affects Intel CPUs dating back a couple generations.

Measurement

Although we have begun the process of changing our power measurements to use an interposer, this dataset uses a current clamp to measure the current at the EPS12V rails. We then multiply that current by the known voltage going into the EPS12V rails to land on the total power consumption in Watts. This does not include total system power, and as a result, allows us to compare even relatively ‘ancient’ data since we are eliminating the impact from GPU components. The motherboard will affect the data (via VRM efficiency losses), but we’re still getting pretty close to the ‘true’ power consumption of the CPU without moving to software measurement, which introduces new issues. The interposer method further improves this, but for purposes of this long-term reference table, we will be presenting and maintaining the clamp method data.

Environment

Testing is conducted in a 21C environment with controls to keep it +/- 1 degree Celsius. Temperature has an impact on power and leakage. All tests since ~2020 are conducted with a 360mm Arctic Liquid Freezer II AIO, with all tests prior to that conducted on a 280mm Liquid Freezer II AIO. Prior to the existence of this Arctic solution, we used the Kraken X62 CPU coolers. In all scenarios, we ensure that the CPU does not thermal throttle. If a CPU is pushing too much power for its cooling, in years past, we’d move to a 360mm CLC from a 280mm solution. Now, all tests are done on a 360 (and that seems to prevent any throttling).

For CPUs wherein a 95C target is part of the design, we let them boost within the limits of their cooling.

Limitations

This testing has limitations, and it’s important that you understand them. We want our audience to be aware of the value offered by the data we present and the limitations of the data.

First, this page currently only presents all-core, 100% load data. That means we are not presenting, for example, the extreme power consumption that might be seen in a Prime95 workload (although Blender is still a heavy, realistic workload). 

Additionally, we are not currently presenting gaming power consumption in these charts (although that is something we’ve recently added, it isn’t public yet). Gaming typically uses less power in total than production workloads due to the more variable nature of gaming impact on a CPU. This becomes more complicated if the bottleneck shifts: If a CPU is fully GPU-bound due to its ultra-high performance, it might appear to draw less power than alternatives that are fully CPU-bound. Again, that’s something we’ll soon be talking about in video format. It’ll make its way to articles later. There are realistic ability and time constraints with a small team, and as such, we focus on presenting the data we find to be the most reliable and which we can maintain with the highest quality. The current approach enables those goals.

Although we collect it, we are also not currently presenting idle power consumption. This is because the idle measurements are often more complicated than stock due to Windows behaviors. We might add these in the future.

CPU Power Consumption - Blender All-Core

The below charts are our latest vetted, full power consumption charts. Data dates back several years. Because it is measured at the EPS12V cables and not total system draw, we are able to provide like-for-like comparisons between these CPUs (sans VRM efficiency losses). Keep in mind that if retested with the newest BIOS, it is possible some of these numbers change.

To keep things legible, we will be breaking these into charts based on generations and architectures that we think people will most commonly compare. The data between all charts should be almost completely cross-comparable, so you can pull two up side-by-side if both parts you're interested in aren't in a single chart.

The axis has been kept at 400W for all charts to make it easier to spot-check comparisons between charts. The one exception is for the Legacy & Misc. chart, which goes to a higher wattage.

Gaming PC Power Supply Wattage - "How Many Watts?"

These CPU power consumption comparison charts will help you understand the full load power requirements of AMD vs. Intel CPUs for each generation. These don't necessarily show power efficiency (that's a different test and requires a unit of work for comparison), but will at least give you a better idea for what kind of power supply you need for your gaming PC build or workstation build. As general guidance, our recommendation for how many watts your PSU needs is to spec for at least maximum power consumption (+ some overhead) to ensure you have enough capacity during peak loads. If you're wondering how big of a power supply to get, start first by looking at benchmarks for CPU and GPU power consumption, then go from there. Although it's unlikely you'd hit 100% load on both components simultaneously, that would allow prep for the worst case scenarios.

AMD Ryzen 7000, Intel 13th Gen, Intel 14th Gen Power Draw

For search reasons and Ctrl+F users, the above chart contains the following CPUs:

AMD Ryzen 5000, Intel 10th, 11th, & 12th Gen Power Consumption

This chart contains the following CPUs:

AMD Ryzen 3000, Intel 8th & 9th Gen Power Consumption

The CPUs in the above chart include:

AMD Ryzen 1000, 2000, & Intel 7th Gen Power Consumption

We don't have as many power consumption numbers dating back to Ryzen 2000, 1000, or the Intel 7th Generation CPUs. We have some random/legacy results (below) that were from various revisits at points in time, but the Ryzen 1000 era and Intel 7th generation would have been when we just started collecting power numbers.

CPUs in the list above include the below, plus various overclocking tests:

Miscellaneous, Older HEDT, & Legacy CPU Power Consumption

The next chart is for miscellaneous and legacy results. These are 'orphaned' tests: We ran them at one point or another, but they don't necessarily belong anywhere else. These results are much more variable based on platform than our newer data. That's because our inventoried platform options were more limited on some of these, like FX.

IMPORTANT: With the oldest CPUs in this table, there is a higher likelihood than typical of a result which does not necessarily match the expectations of the time. The result would be 'accurate' in the sense that the current clamp would have been accurately used, but the BIOS settings may not match whatever the "guidance" was from manufacturers at the time, as we were not receiving guidance in the pre-2015 era for AMD or Intel. In other words, we would have exercised less control over BIOS as we did not have official guidance on the expectations (since we were not part of the reviewer circuit).

The above chart contains these CPUs:

Update Log

The below contains an update log of changes to this page. The format is MM/DD/YYYY:

• 12/03/2023: Added table of contents feature to ease navigation
• 11/30/2023: Created page with initial dataset following Threadripper 7000 reviews

A living document detailing the current and retired test platforms used by GamersNexus

The Highlights

• This document will be updated as we add new benches
• Old test benches will be kept on this page, but moved to the bottom
• Remember that testing regularly requires changes to test platforms to accommodate the DUT

Table of Contents

• AutoTOC

Intro

This page is a living document that will be updated as we revolve test platforms through. The page will be organized by type of test bench and year. We regularly cycle benches into retirement as we update for new components. We have had numerous manufacturers inform us that they try to keep tabs on our bench platforms for their own internal research and performance analysis, so we hope that this may help manufacturers with their product development when attempting to isolate performance concerns we present.

And to reiterate: This page is NOT COMPLETE. This is under construction. There is a lot of information still missing here.

Latest Update (January 17, 2025): We have added several test platforms to this page, including our newest GPU test bench.

Credits

Test Lead, Writing

Steve Burke

Writing

Patrick Lathan

Objectives

This page will primarily serve as a landing page for test benches that we commonly use. This page is still being built and is under construction. The objective is to provide a single stop to answer questions about what hardware we use.

Our philosophy is to review each component in a vacuum, eliminating as many variables and bottlenecks imposed by other components as possible. This means that GPU testing for GPU performance will use a relatively high-powered CPU and RAM, thus hopefully limiting bottlenecks (to the extent possible) outside of the GPU; however, for that same GPU review, we would use a different test bench for acoustics. That's because the acoustics testing does not require unbound performance to eliminate variables, but instead requires a total (within reason) elimination of noise from non-GPU components instead. In this example, we would use the GPU test bench for thermals and gaming, but the GPU acoustic bench for noise levels and frequency response. Likewise, we use a different bench entirely for GPU power consumption measurements, as the host platform is irrelevant for these and only the power consumption matters (and it allows us to asynchronously test power consumption while our primary benches remain occupied for testing gaming or thermals, which greatly improves logistics to allow more testing in a shorter time period).

A lot of this has to do with a general internal approach of specializing within the team. Typically, the same person will handle GPU power consumption tests for every review, but a different person will handle GPU gaming performance and analysis, and a different person still might handle acoustics. Splitting benches like this not only isolates for variables, but also allows the team's experts in each test method to operate without blocking their teammates, as the benches are not singularly needed for every test in a suite.

Caveats & Limitations

We sometimes have to change-out single components (or revert to older benches) for certain tests. For example, if we are running a suite of CPU tests that requires us to go back to AM4 or Z270 or similar, we obviously won't be using the standardized Intel or AMD platform listed as the current primary, but would instead use one of the prior platforms. We will try to keep the older (retired) platforms listed further down the page for such events, but it will take us some time to comb through the site history and catalog them all in this format.

These test benches might not perfectly represent each individual content piece. Some pieces require a singular variable change to these benches as part of the piece. For example, if we're testing something like Intel APO, we will use a different boot drive (to avoid cross-contamination with the primary bench drive) and will use different driver packages and a potentially different BIOS configuration (to enable the dynamic tuning). Likewise, if we're testing PCIe bandwidth scaling and impact on an NVIDIA *90-class card, we will change the PCIe slot's generation for that test. The DUT (Device Under Test) or feature under test will deviate from the fixed bench as part of the test.

Sometimes we miss listing a detail since there are so many, but we will do our best to explain these caveats in each individual content piece.

Consistency is What Matters

Consistency is all that really matters. For comparative reviews, which is how we conduct our reviews, we just need an unchanging and fixed platform that allows us to reliably see scaling and differences between components. In the above example then, having APO toggling or PCIe generation switches being the new variable on an otherwise unchanging platform allows us the most comparative data against our "baseline" for that platform.

Likewise, this means that components being "old" or back by a couple generations becomes largely irrelevant. It only starts mattering if it's meaningfully limiting either the performance scaling of the product or the technology that can be used for the product. For instance, for purposes of generating heat for a cooler review, the CPU itself doesn't matter as much as its power consumption (and IHS shape, to be fair).

Methodology Pieces

This contains a list of our recently published or current test methodologies. Note that we have not yet gone back to upfit all of the older pieces for this website revision, so some links are to the videos.

This section is under construction. We still have to build this section out fully. It is missing a lot of information.

Recent Game Test Methodologies & Research

• Dragon's Dogma 2 (CPU & GPU Research)
• Starfield: (GPU Research | CPU Research | Graphics Optimization Guide) (Videos)
• Cities Skylines 2: GPU Research
• Baldur's Gate 3: GPU Research

Hardware Test Methodologies & Research

• CPU Cooler Test Methodology (Article) and accompanying video
• GPU Test Methodology (basics - the games have changed, some concepts are the same)

Test Benches

Mini-ITX Case Variable Test Bench (2023-TBD)

GPU Acoustics Test Bench (2023-TBD)

The below test bench details our platform used for GPU acoustics. This is deployed for collecting data such as the below. The GPU Acoustics Bench is not used for performance analysis (other than acoustic performance), and so the core components do not matter. For this bench, the most important aspect is controlling the noise: We use passive cooling, like from the Noctua NH-P1, to ensure no noise interference from other parts of the system. The power supply is used in a 0-RPM mode and validated during testing. The GPU does not perform work, but rather has its fan speed manually adjusted to meet the automatic fan speed we find during load testing (from our GPU FPS & thermal test platform, also on this page). We also adjust the speed in increments of 5% to capture a dBA profile across the RPM range.

Unless otherwise stated, video card coolers are tested at an industry-standard distance of 1m from the center fan (aligned on the horizontal and vertical center of the video card cooler). The video card is situated in a horizontal Thermaltake Core P3 platform with the glass removed, so it is basically just a flat "case."

GPU Test Benches

The below section includes current and deprecated GPU test benches. These are benches designed to isolate the GPU as much as possible for purposes of benchmarking and review.

Primary GPU Test Bench (2024-Present)

The below test bench is our current GPU test and review platform.

We have three of these that are calibrated against each other. We keep a data set for occasional calibration to ensure there isn't drift. At least one GPU is rerun for each cycle to check for drift, game patch impact, Windows impact, or other changes and how they may affect results. If we find deviation from the most recent data set, we re-run everything. Dates are logged for each device tested in the charts.

Two of these are used for GPU reviews and benchmarks, with the third used for power and efficiency testing.

GamersNexus has bought all components for its GPU test benches at this point, with the exception of the remnant SuperNOVA PSUs from years ago. They're still running great!

Additional parameters include:

• Hardware-accelerated GPU scheduling ENABLED
• ReBAR ENABLED
• Power plan set to High Performance
• Power testing uses identical bench with calibrated Elmor PMD2

Secondary GPU Test Bench (2025-TBD)

The below test bench is a special configuration for occasional focused testing on lower-end CPUs. This is not used for every review cycle, but as of the Intel Arc B570 review, we introduced it for limited run testing on lower-end hardware. We will likely not use this platform for high-end GPUs and will only introduce it for cards under the $300 mark (and even then, only occasionally -- it is a tripling of work to maintain).

5600X Platform (CPU-Limited GPU Benchmarks)

12400 Platform (CPU-Limited GPU Benchmarks)

GPU Games Suite (2024-Present)

GPU Test Bench (2022-2023)

The below test bench details our two prior GPU review platforms that were used for performance evaluation and thermals. These deprecated and retired GPU test benches were overclocked, which helped significantly limit the bottlenecks potentially imposed by a CPU. The platforms also ran tests at 1080p, 1440p, and 4K, with the higher resolutions further distancing us from CPU scaling limitations. At EOL, this platform was primarily GPU-bound, but 1080p was starting to become constrained for high-end cards. This platform was retired from service at the end of 2024.

Test Applications & Games (DEPRECATED)

We are constantly changing the games that get tested on this platform, and as such, maintaining a list of what we currently show is difficult. Instead, we'll list everything that the platform runs as of November 2, 2023. This list will likely be moved to the GPU methodology piece once we get it into the new website.

Additional tests are run, but are not explicitly listed as they may be more variable (or we just haven't publicly documented them at this time).

CPU Cooler Test Bench (2020-2024)

CPU Test Bench (2024-TBD)

We are still working to fully add our modernized testing methodology for CPU benchmarks and reviews. Often, the review itself will contain verbal explanations of each test and methodological notes. We are slowly compiling those as we continue to build-out the living doc on the website. The below tests are typically run on every CPU; however, we sometimes remove tests for reasons that could relate to time with the product before embargo lift, compatibility, performance being too low (e.g. a CPU that struggles to maintain 30 FPS might just see that test eliminated), game updates that invalidate prior data (and need to be recollected), or if we believe the charts in a review are enough to illustrate the pattern without lengthening the review further.

Our CPU tests include these games and settings:

Production Workloads

Gaming Workloads

CPU Miscellaneous Tests

We will add the test bench hardware information for the next round of reviews, but the vast majority of critical information is listed in the chart subtitles.

Check back soon: We still need to add our case bench, our new cooler bench, and our CPU benches, among other platforms!

Update Log

• January 17, 2025: Added modern GPU test platform. Moved prior platform to 'deprecated' status. Added CPU-limited GPU test platforms.
• January 17, 2025: Update Log created. Although there were prior updates, we did not log them here.