AMD

Review: New Framework Laptop 16 takes a fresh stab at the upgradeable laptop GPU

AMD, AMD Ryzen, Features, framework, framework laptop 13, framework laptop 16, gaming, NVIDIA, nvidia geforce, Reviews, right to repair, Tech / 9u50fv / November 12, 2025

framework laptop 16, take two

New components make it more useful and powerful but no less odd.

Credit: Andrew Cunningham

The original Framework Laptop 16 was trying to crack a problem that laptop makers have wrestled with on and off for years: Can you deliver a reasonably powerful, portable workstation and gaming laptop that supports graphics card upgrades just like a desktop PC?

Specs at a glance: Framework Laptop 16 (2025)
OS	Windows 11 25H2
CPU	AMD Ryzen AI 7 350 (4 Zen 5 cores, 4 Zen 5c cores)
RAM	32GB DDR5-5600 (upgradeable)
GPU	AMD Radeon 860M (integrated)/Nvidia GeForce RTX 5070 Mobile (dedicated)
SSD	1TB Western Digital Black SN770
Battery	85 WHr
Display	16-inch 2560×1600 165 Hz matte non-touchscreen
Connectivity	6x recessed USB-C ports (2x USB 4, 4x USB 3.2) with customizable “Expansion Card” dongles
Weight	4.63 pounds (2.1 kg) without GPU, 5.29 pounds (2.4 kg) with GPU
Price as tested	Roughly $2,649 for pre-built edition; $2,517 for DIY edition with no OS

Even in these days of mostly incremental, not-too-exciting GPU upgrades, the graphics card in a gaming PC or graphics-centric workstation will still feel its age faster than your CPU will. And the chance to upgrade that one component for hundreds of dollars instead of spending thousands replacing the entire machine is an appealing proposition.

Upgradeable, swappable GPUs would also make your laptop more flexible—you can pick and choose from various GPUs from multiple vendors based on what you want and need, whether that’s raw performance, power efficiency, Linux support, or CUDA capabilities.

Framework’s first upgrade to the Laptop 16—the company’s first upgrade to any of its products aside from the original Laptop 13—gets us pretty close to that reality. The laptop can now support two interchangeable motherboards: one with an older AMD Ryzen 7040-series CPU and one with a new Ryzen AI 300-series CPU. And both motherboards can be used either with just an integrated GPU or with dedicated GPUs from both AMD and Nvidia.

The Nvidia GeForce 5070 graphics module is the most exciting and significant part of this batch of updates, but there are plenty of other updates and revisions to the laptop’s external and internal components, too. These upgrades don’t address all of our problems with the initial version of the laptop, but they do help quite a bit. And a steady flow of updates like these would definitely make the Laptop 16 a platform worth investing in.

Re-meet the Framework Laptop 16

Framework’s Laptop 13 stacked on top of the 16. Credit: Andrew Cunningham

Framework treats each of its laptops as a platform to be modified and built upon rather than something to be wholly redesigned and replaced every time it’s updated. So these reviews necessarily re-cover ground we have already covered—I’ve also reused some of the photos from last time, since this is quite literally the same laptop in most respects. I’ll point you to the earlier review for detailed notes on the build process and how the laptop is put together.

To summarize our high-level notes about the look, feel, and design of the Framework Laptop 16: While the Framework Laptop 13 can plausibly claim to be in the same size and weight class as portables like the 13-inch MacBook Air, the Framework Laptop 16 is generally larger and heavier than the likes of the 16-inch MacBook Pro or portable PC workstations like the Lenovo ThinkPad P1 or Dell 16 Premium. That’s doubly true once you actually add a dedicated graphics module to the Laptop 16—these protrude a couple of inches from the back of the laptop and add around two-thirds of a pound to its weight.

	Frame-work 16 (no GPU)	Frame-work 16 (GPU)	Apple 16-inch MBP	Dell 16 Premium	Lenovo ThinkPad P1 Gen 8	HP ZBook X G1i	Lenovo Legion Pro 5i Gen 10	Razer Blade 16
Size (H x W x D inches)	0.71 x 14.04 x 10.63	0.82 x 14.04 x 11.43	0.66 x 14.01 x 9.77	0.75 x 14.1 x 9.4	0.39-0.62 x 13.95 x 9.49	0.9 x 14.02 x 9.88	0.85-1.01 x 14.34 x 10.55	0.59-0.69 x 13.98 x 9.86
Weight	4.63 lbs	5.29 lbs	4.7-4.8 lbs	4.65 pounds	4.06 lbs	4.5 lbs	5.56 lbs	4.71 lbs

You certainly can find laptops from the major PC OEMs that come close to or even exceed the size and weight of the Laptop 16. But in most cases, you’ll find that comparably specced and priced laptops are an inch or two less deep and at least half a pound lighter than the Laptop 16 with a dedicated GPU installed.

But if you’re buying from Framework, you’re probably at least notionally interested in customizing, upgrading, and repairing your laptop over time, all things that Framework continues to do better than any other company.

The Laptop 16’s customizable keyboard deck is still probably its coolest feature—it’s a magnetically attached series of panels that allows you to remove and replace components without worrying about the delicate and finicky ribbon cables the Laptop 13 uses. Practically, the most important aspect of this customizable keyboard area is that it lets you decide whether you want to install a dedicated number pad or not; this also allows you to choose whether you want the trackpad to be aligned with the center of the laptop or with wherever the middle of the keyboard is.

It might look a little rough, but the customizable keyboard deck is still probably the coolest thing about the Laptop 16 in day-to-day use. Andrew Cunningham

But Framework also sells an assortment of other functional and cosmetic panels and spacers to let users customize the laptop to their liking. The coolest, oddest accessories are still probably the LED matrix spacers and the clear, legend-less keyboard and number pad modules. We still think this assortment of panels gives the system a vaguely unfinished look, but Framework is clearly going for function over form here.

The Laptop 16 also continues to use Framework’s customizable, swappable Expansion Card modules. In theory, these let you pick the number and type of ports your laptop has, as well as customize your port setup on the fly based on what you need. But as with all AMD Ryzen-based Framework Laptops, there are some limits to what each port can do.

According to Framework’s support page, there’s no single Expansion Card slot that is truly universal:

Ports 1 and 4 support full 40Gbps USB 4 transfer speeds, display outputs, and up to 240 W charging, but if you use a USB-A Expansion Card in those slots, you’ll increase power use and reduce battery life.
Ports 2 and 4 support display outputs, up to 240 W charging, and lower power usage for USB-A ports, but they top out at 10Gbps USB 3.2 transfer speeds. Additionally, port 5 (the middle port on the right side of the laptop, if you’re looking at it head-on) supports the DisplayPort 1.4 standard where the others support DisplayPort 2.1.
Ports 3 and 4 are limited to 10Gbps USB 3.2 transfer speeds and don’t support display outputs or charging.

The Laptop 16 also doesn’t include a dedicated headphone jack, so users will need to burn one of their Expansion Card slots to get one.

Practically speaking, most users will be able to come up with a port arrangement that fits their needs, and it’s still handy to be able to add and remove things like Ethernet ports, HDMI ports, or SD card readers on an as-needed basis. But choosing the right Expansion Card slot for the job will still require some forethought, and customizable ports aren’t as much of a selling point for a 16-inch laptop as they are for a 13-inch laptop (the Framework Laptop 13 was partly a response to laptops like the MacBook Air and Dell XPS 13 that only came with a small number of USB-C ports; larger laptops have mostly kept their larger number and variety of ports).

What’s new in 2025’s Framework Laptop 16?

An upgraded motherboard and a new graphics module form the heart of this year’s Laptop 16 upgrade. The motherboard steps up from AMD Ryzen 7040-series processors to AMD Ryzen AI 7 350 and Ryzen AI 9 HX 370 chips. These are the same processors Framework put into the Laptop 13 earlier this year, though they ought to be able to run a bit faster in the Laptop 16 due to its larger heatsink and dual-fan cooling system.

Along with an upgrade from Zen 4-based CPU cores to Zen 5 cores, the Ryzen AI series includes an upgraded neural processing unit (NPU) that is fast enough to earn Microsoft’s Copilot+ PC label. These PCs have access to a handful of unique Windows 11 AI and machine-learning features (yes, Recall, but not just Recall) that are processed locally rather than in the cloud. If you don’t care about these features, you can mostly just ignore them, but if you do care, this is the first version of the Laptop 16 to support them.

Most of the new motherboard’s other specs and features are pretty similar to the first-generation version; there are two SO-DIMM slots for up to 96GB of DDR5-5600, one M.2 2280 slot for the system’s main SSD, and one M.2 2230 slot for a secondary SSD. Wi-Fi 7 and Bluetooth connectivity are provided by an AMD RZ717 Wi-Fi card that can at least theoretically also be replaced with something faster down the line if you want.

The more exciting upgrade, however, may be the GeForce RTX 5070 GPU. This is the first time Framework has offered an Nvidia product—its other GPUs have all come from either Intel or AMD—and it gives the new Laptop 16 access to Nvidia technologies like DLSS and CUDA, as well as much-improved performance for games with ray-traced lighting effects.

Those hoping for truly high-end graphics options for the Laptop 16 will need to keep waiting, though. The laptop version of the RTX 5070 is actually the same chip as the desktop version of the RTX 5060, a $300 graphics card with 8GB of RAM. As much as it adds to the Laptop 16, it still won’t let you come anywhere near 4K in most modern games, and for some, it may even struggle to take full advantage of the internal 165 Hz 1600p screen. Professional workloads (including AI workloads) that require more graphics RAM will also find the mobile 5070 lacking.

Old 180 W charger on top, new 240 W charger on bottom. Credit: Andrew Cunningham

Other components have gotten small updates as well. For those who upgrade an existing Laptop 16 with the new motherboard, Framework is selling 2nd-generation keyboard and number pad components. But their main update over the originals is new firmware that “includes a fix to prevent the system from waking while carried in a bag.” Owners of the original keyboard can install a firmware update to get the same functionality (and make their input modules compatible with the new board).

Upgraders should also note that the original system’s 180 W power adapter has been replaced with a 240 W model, the maximum amount of power that current USB-C and USB-PD standards are capable of delivering. You can charge the laptop with just about any USB-C power brick, but anything lower than 240 W risks reducing performance (or having the battery drain faster than it can charge).

Finally, the laptop uses a second-generation 16-inch, 2560×1600, 165 Hz LCD screen. It’s essentially identical in every way to the first-generation screen, but it formally supports G-Sync, Nvidia’s adaptive sync implementation. The original screen can still be used with the new motherboard, but it only supports AMD’s FreeSync, and Framework told us a few months ago that the panel supplier had no experience providing consumer-facing firmware updates that might add G-Sync to the old display. It’s probably not worth replacing the entire screen for, but it’s worth noting whether you’re upgrading the laptop or buying a new one.

Performance

Framework sent us the lower-end Ryzen AI 7 350 processor configuration for our new board, making it difficult to do straightforward apples-to-apples comparisons to the high-end Ryzen 9 7940HS in our first-generation Framework board. We did test the new chip, and you’ll see its results in our charts.

We’ve also provided numbers from the Ryzen AI 9 HX 370 in the Asus Zenbook S16 UM5606W to show approximately where you can expect the high-end Framework Laptop 16 configuration to land (Framework’s integrated graphics performance will be marginally worse since it’s using slower socketed RAM rather than LPDDR5X; other numbers may differ based on how each manufacturer has configured the chip’s power usage and thermal behavior). We’ve also included numbers from the same chip in the Framework Laptop 13, though Framework’s spec sheets indicate that the chips have different power limits and thus will perform differently.

We were able to test the new GeForce GPU in multiple configurations—both paired with the new Ryzen AI 7 350 processor and with the old Ryzen 9 7940HS chip. This should give anyone who bought the original Laptop 16 an idea of what kind of performance increase they can expect from the new GPU alone. In all, we’ve tested or re-tested:

The Ryzen 7 7940HS CPU from the first-generation Laptop 16 and its integrated Radeon 780M GPU
The Ryzen 7 7940HS and the original Radeon RX 7700S GPU module
The Ryzen 7 7940HS and the new GeForce RTX 5070 GPU module, for upgraders who only want to grab the new GPU
The Ryzen AI 7 350 CPU and the GeForce RTX 5070 GPU

We also did some light testing on the Radeon 860M integrated GPU included with the Ryzen AI 7 350.

All the Laptop 16 performance tests were run with Windows’ Best Performance power preset enabled, which will slightly boost performance at the expense of power efficiency.

Given all of those hardware combinations, we simply ran out of time to test the new motherboard with the old Radeon RX 7700S GPU—Framework is continuing to sell it, so it is a realistic combination of components. But our RTX 5070 testing suggests that these GPUs will perform pretty much the same regardless of which CPU you pair them with.

If you’re buying the cheaper Laptop 16 with the Ryzen AI 7 350, the good news is that it generally performs at least as well as—and usually a bit better than—the high-end Ryzen 9 7940HS from the last-generation model. Performance is also pretty similar to the Ryzen AI 9 HX 370 in smaller, thinner laptops—the extra power and cooling capacity in the Laptop 16 is paying off here. People choosing between a PC and a Mac should note that none of these Ryzen chips come anywhere near the M4 Pro used in comparably priced 16-inch MacBook Pros, but that’s just where the PC ecosystem is these days.

How big an upgrade the GeForce 5070 will be depends on the game you’re playing. In titles like Borderlands 3 that naturally run a bit better on AMD’s GPUs, there’s not much of a difference at all. In games like Cyberpunk 2077 with heavy ray-tracing effects enabled, the mobile RTX 5070 can be nearly twice as fast as the RX 7700S.

Most games will fall somewhere in between those two extremes; our tests show that the improvements hover between 20 and 30 percent most of the time, just a shade less than the 30 to 40 percent improvement that Framework claimed in its original announcement.

Beyond raw performance, the other thing you get with an Nvidia GPU is access to a bunch of important proprietary technologies like DLSS upscaling and CUDA—these technologies are often better and more widely supported than the equivalent technologies that AMD’s or Intel’s GPUs use, thanks in part to Nvidia’s overall dominance of the dedicated GPU market.

In the tests we’ve run on them, the Radeon 860M and 890M are both respectable integrated GPUs (the lower-end 860M typically falls just short of last generation’s top-end 780M, but it’s very close). They’re never able to provide more than a fraction of the Radeon RX 7700S’s performance, let alone the RTX 5070, but they’ll handle a lot of lighter games at 1080p. I would not buy a system this large or heavy just to use it with an integrated GPU.

Better to be unique than perfect

It’s expensive and quirky, but the Framework Laptop 16 is worth considering because it’s so different from what most other laptop makers are doing. Credit: Andrew Cunningham

Our original Framework Laptop 16 review called it “fascinating but flawed,” and the parts that made it flawed haven’t really changed much over the last two years. It’s still relatively large and heavy; the Expansion Card system still makes less sense in a larger laptop than it does in a thin-and-light; the puzzle-like grid of input modules and spacers looks kind of rough and unfinished.

But the upgrades do help to shift things in the Laptop 16’s favor. Its modular and upgradeable design was always a theoretical selling point, but the laptop now actually offers options that other laptops don’t.

The presence of both AMD and Nvidia GPUs is a big step up in flexibility for both gaming and professional applications. The GeForce module is a better all-around choice, with slightly to significantly faster game performance and proprietary technologies like DLSS and CUDA, while the Radeon GPU is a cheaper option with better support for Linux.

Given their cost, I still wish that these GPUs were more powerful—they’re between $350 or $449 for the Radeon RX 7700S and between $650 and $699 for the RTX 5070 (prices vary a bit and are cheaper when you’re buying them together with a new laptop rather than buying them separately). You’ll basically always spend more for a gaming laptop than you will for a gaming desktop with similar or better performance, but that does feel like an awful lot to spend for GPUs that are still limited to 8GB of RAM.

Cost is a major issue for the Laptop 16 in general. You may save money in the long run by buying a laptop that you can replace piece-by-piece as you need to rather than all at once. But it’s not even remotely difficult to find similar specs from the major PC makers for hundreds of dollars less. We can’t vouch for the build quality or longevity of any of those PCs, but it does mean that you have to be willing to pay an awful lot just for Framework’s modularity and upgradeability. That’s true to some degree of the Laptop 13 as well, but the price gap between the 13 and competing systems isn’t as large as it is for the 16.

Whatever its lingering issues, the Framework Laptop 16 is still worth considering because there’s nothing else quite like it, at least if you’re in the market for something semi-portable and semi-powerful. The MacBook Pro exists if you want something more appliance-like, and there’s a whole spectrum of gaming and workstation PCs in between with all kinds of specs, sizes, and prices. To stand out from those devices, it’s probably better to be unique than to be perfect, and the reformulated Laptop 16 certainly clears that bar.

The good

Modular, repairable, upgradeable design that’s made to last
Cool, customizable keyboard deck
Nvidia GeForce GPU option gives the Laptop 16 access to some gaming and GPU computing features that weren’t usable with AMD GPUs
GPU upgrade can be added to first-generation Framework Laptop 16
New processors are a decent performance improvement and are worth considering for new buyers
Old Ryzen 7040-series motherboard is sticking around as an entry-level option, knocking $100 off the former base price ($1,299 and up for a barebones DIY edition, $1,599 and up for the cheapest pre-built)
Framework’s software support has gotten better in the last year

The bad

Big and bulky for the specs you get
Mix-and-match input modules and spacers give it a rough, unfinished sort of look
Ryzen AI motherboards are more expensive than the originals were when they launched

The ugly

It’ll cost you—the absolute bare minimum price for Ryzen AI 7 350 and RTX 5070 combo is $2,149, and that’s without RAM, an SSD, or an operating system

Andrew is a Senior Technology Reporter at Ars Technica, with a focus on consumer tech including computer hardware and in-depth reviews of operating systems like Windows and macOS. Andrew lives in Philadelphia and co-hosts a weekly book podcast called Overdue.

Review: New Framework Laptop 16 takes a fresh stab at the upgradeable laptop GPU Read More »

After confusing driver release, AMD says old GPUs are still actively supported

AMD, gaming, Radeon, rdna, rdna 2, Tech / Shannon Garcia / November 4, 2025

The release notes for the 25.10.2 Adrenalin release also dropped Windows 10 from the list of “compatible operating systems,” listing only Windows 11 21H2 and later. But AMD confirmed to Windows Latest that the driver packages would still support Windows 10 for the foreseeable future. The company said that the OS is not listed in the release notes because Microsoft has technically ended support for Windows 10, but home users running Windows 10 on their PCs can get an extra year of security patches relatively easily. Microsoft will continue to provide support for the OS in businesses, schools, and other large organizations until at least 2028.

Why all the fuss?

It would look bad if AMD dropped or reduced support for those Radeon 5000- and 6000-series GPUs, given that Nvidia continues to support GeForce RTX 20- and 30-series graphics cards launched in the same 2019 to 2022 time window. But the end of support could have been even worse for gaming handhelds and lower-end PCs with integrated graphics.

The RDNA 2 architecture, in particular, has enjoyed a long and ongoing life as an integrated GPU, including for systems that are explicitly marketed and sold as gaming PCs. And because so many of AMD and Intel’s lower-end chips are just rebranded versions of older silicon, AMD continues to launch “new” products with RDNA 2 GPUs. The RDNA 2 architecture is the one Valve has used in the Steam Deck since 2022, for example, but Microsoft and Asus’ just-launched ROG Xbox Ally series also includes an RDNA 2 GPU in the entry-level model.

The last time AMD formally scaled back its GPU driver support was in 2023, when it moved drivers for its Polaris and Vega GPU architectures into a separate package that would only get occasional “critical updates.” At the time, AMD had launched its last dedicated Vega-based GPU just four years before, and many lower-end desktop and laptop processors still shipped with Vega-based integrated GPUs.

For the Steam Deck and other SteamOS and Linux systems, at least, it seems that things aren’t really changing, no matter what happens with the Windows drivers. Phoronix points out that the Linux driver package for AMD’s GPUs has always been maintained separately from the Windows drivers and that GPU architectures considerably older than RDNA 1 continue to get official support and occasional improvements.

After confusing driver release, AMD says old GPUs are still actively supported Read More »

New physical attacks are quickly diluting secure enclave defenses from Nvidia, AMD, and Intel

AMD, battering ram, Biz & IT, Features, intel, NVIDIA, secure enclaves, Security, tee.fail, trusted execution environments, wiretap / Shannon Garcia / October 29, 2025

On-chip TEEs withstand rooted OSes but fall instantly to cheap physical attacks.

Trusted execution environments, or TEEs, are everywhere—in blockchain architectures, virtually every cloud service, and computing involving AI, finance, and defense contractors. It’s hard to overstate the reliance that entire industries have on three TEEs in particular: Confidential Compute from Nvidia, SEV-SNP from AMD, and SGX and TDX from Intel. All three come with assurances that confidential data and sensitive computing can’t be viewed or altered, even if a server has suffered a complete compromise of the operating kernel.

A trio of novel physical attacks raises new questions about the true security offered by these TEES and the exaggerated promises and misconceptions coming from the big and small players using them.

The most recent attack, released Tuesday, is known as TEE.fail. It defeats the latest TEE protections from all three chipmakers. The low-cost, low-complexity attack works by placing a small piece of hardware between a single physical memory chip and the motherboard slot it plugs into. It also requires the attacker to compromise the operating system kernel. Once this three-minute attack is completed, Confidential Compute, SEV-SNP, and TDX/SDX can no longer be trusted. Unlike the Battering RAM and Wiretap attacks from last month—which worked only against CPUs using DDR4 memory—TEE.fail works against DDR5, allowing them to work against the latest TEEs.

Some terms apply

All three chipmakers exclude physical attacks from threat models for their TEEs, also known as secure enclaves. Instead, assurances are limited to protecting data and execution from viewing or tampering, even when the kernel OS running the processor has been compromised. None of the chipmakers make these carveouts prominent, and they sometimes provide confusing statements about the TEE protections offered.

Many users of these TEEs make public assertions about the protections that are flat-out wrong, misleading, or unclear. All three chipmakers and many TEE users focus on the suitability of the enclaves for protecting servers on a network edge, which are often located in remote locations, where physical access is a top threat.

“These features keep getting broken, but that doesn’t stop vendors from selling them for these use cases—and people keep believing them and spending time using them,” said HD Moore, a security researcher and the founder and CEO of runZero.

He continued:

Overall, it’s hard for a customer to know what they are getting when they buy confidential computing in the cloud. For on-premise deployments, it may not be obvious that physical attacks (including side channels) are specifically out of scope. This research shows that server-side TEEs are not effective against physical attacks, and even more surprising, Intel and AMD consider these out of scope. If you were expecting TEEs to provide private computing in untrusted data centers, these attacks should change your mind.

Those making these statements run the gamut from cloud providers to AI engines, blockchain platforms, and even the chipmakers themselves. Here are some examples:

Cloudflare says it’s using Secure Memory Encryption—the encryption engine driving SEV—to safeguard confidential data from being extracted from a server if it’s stolen.
In a post outlining the possibility of using the TEEs to secure confidential information discussed in chat sessions, Anthropic says the enclave “includes protections against physical attacks.”
Microsoft marketing (here and here) devotes plenty of ink to discussing TEE protections without ever noting the exclusion.
Meta, paraphrasing the Confidential Computing Consortium, says TEE security provides protections against malicious “system administrators, the infrastructure owner, or anyone else with physical access to the hardware.” SEV-SNP is a key pillar supporting the security of Meta’s WhatsApp Messenger.
Even Nvidia claims that its TEE security protects against “infrastructure owners such as cloud providers, or anyone with physical access to the servers.”
The maker of the Signal private messenger assures users that its use of SGX means that “keys associated with this encryption never leave the underlying CPU, so they’re not accessible to the server owners or anyone else with access to server infrastructure.” Signal has long relied on SGX to protect contact-discovery data.

I counted more than a dozen other organizations providing assurances that were similarly confusing, misleading, or false. Even Moore—a security veteran with more than three decades of experience—told me: “The surprising part to me is that Intel/AMD would blanket-state that physical access is somehow out of scope when it’s the entire point.”

In fairness, some TEE users build additional protections on top of the TEEs provided out of the box. Meta, for example, said in an email that the WhatsApp implementation of SEV-SNP uses protections that would block TEE.fail attackers from impersonating its servers. The company didn’t dispute that TEE.fail could nonetheless pull secrets from the AMD TEE.

The Cloudflare theft protection, meanwhile, relies on SME—the engine driving SEV-SNP encryption. The researchers didn’t directly test SME against TEE.fail. They did note that SME uses deterministic encryption, the cryptographic property that causes all three TEEs to fail. (More about the role of deterministic encryption later.)

Others who misstate the TEEs’ protections provide more accurate descriptions elsewhere. Given all the conflicting information, it’s no wonder there’s confusion.

How do you know where the server is? You don’t.

Many TEE users run their infrastructure inside cloud providers such as AWS, Azure, or Google, where protections against supply-chain and physical attacks are extremely robust. That raises the bar for a TEE.fail-style attack significantly. (Whether the services could be compelled by governments with valid subpoenas to attack their own TEE is not clear.)

All these caveats notwithstanding, there’s often (1) little discussion of the growing viability of cheap, physical attacks, (2) no evidence (yet) that implementations not vulnerable to the three attacks won’t fall to follow-on research, or (3) no way for parties relying on TEEs to know where the servers are running and whether they’re free from physical compromise.

“We don’t know where the hardware is,” Daniel Genkin, one of the researchers behind both TEE.fail and Wiretap, said in an interview. “From a user perspective, I don’t even have a way to verify where the server is. Therefore, I have no way to verify if it’s in a reputable facility or an attacker’s basement.”

In other words, parties relying on attestations from servers in the cloud are once again reduced to simply trusting other people’s computers. As Moore observed, solving that problem is precisely the reason TEEs exist.

In at least two cases, involving the blockchain services Secret Network and Crust, the loss of TEE protections made it possible for any untrusted user to present cryptographic attestations. Both platforms used the attestations to verify that a blockchain node operated by one user couldn’t tamper with the execution or data passing to another user’s nodes. The Wiretap hack on SGX made it possible for users to run the sensitive data and executions outside of the TEE altogether while still providing attestations to the contrary. In the AMD attack, the attacker could decrypt the traffic passing through the TEE.

Both Secret Network and Crust added mitigations after learning of the possible physical attacks with Wiretap and Battering RAM. Given the lack of clear messaging, other TEE users are likely making similar mistakes.

A predetermined weakness

The root cause of all three physical attacks is the choice of deterministic encryption. This form of encryption produces the same ciphertext each time the same plaintext is encrypted with the same key. A TEE.fail attacker can copy ciphertext strings and use them in replay attacks. (Probabilistic encryption, by contrast, resists such attacks because the same plaintext can encrypt to a wide range of ciphertexts that are randomly chosen during the encryption process.)

TEE.fail works not only against SGX but also a more advanced Intel TEE known as TDX. The attack also defeats the protections provided by the latest Nvidia Confidential Compute and AMD SEV-SNP TEEs. Attacks against TDX and SGX can extract the Attestation Key, an ECDSA secret that certifies to a remote party that it’s running up-to-date software and can’t expose data or execution running inside the enclave. This Attestation Key is in turn signed by an Intel X.509 digital certificate providing cryptographic assurances that the ECDSA key can be trusted. TEE.fail works against all Intel CPUs currently supporting TDX and SDX.

With possession of the key, the attacker can use the compromised server to peer into data or tamper with the code flowing through the enclave and send the relying party an assurance that the device is secure. With this key, even CPUs built by other chipmakers can send an attestation that the hardware is protected by the Intel TEEs.

GPUs equipped with Nvidia Confidential Compute don’t bind attestation reports to the specific virtual machine protected by a specific GPU. TEE.fail exploits this weakness by “borrowing” a valid attestation report from a GPU run by the attacker and using it to impersonate the GPU running Confidential Compute. The protection is available on Nvidia’s H100/200 and B100/200 server GPUs.

“This means that we can convince users that their applications (think private chats with LLMs or Large Language Models) are being protected inside the GPU’s TEE while in fact it is running in the clear,” the researchers wrote on a website detailing the attack. “As the attestation report is ‘borrowed,’ we don’t even own a GPU to begin with.”

SEV-SNP (Secure Encrypted Virtualization-Secure Nested Paging) uses ciphertext hiding in AMD’s EPYC CPUs based on the Zen 5 architecture. AMD added it to prevent a previous attack known as Cipherleaks, which allowed malicious hypervisors to extract cryptographic keys stored in the enclaves of a virtual machine. Ciphertext, however, doesn’t stop physical attacks. With the ability to reopen the side channel that Cipherleaks relies on, TEE.fail can steal OpenSSL credentials and other key material based on constant-time encryption.

Cheap, quick, and the size of a briefcase

“Now that we have interpositioned DDR5 traffic, our work shows that even the most modern of TEEs across all vendors with available hardware is vulnerable to cheap physical attacks,” Genkin said.

The equipment required by TEE.fail runs off-the-shelf gear that costs less than $1,000. One of the devices the researchers built fits into a 17-inch briefcase, so it can be smuggled into a facility housing a TEE-protected server. Once the physical attack is performed, the device does not need to be connected again. Attackers breaking TEEs on servers they operate have no need for stealth, allowing them to use a larger device, which the researchers also built.

A logic analyzer attached to an interposer.

The researchers demonstrated attacks against an array of services that rely on the chipmakers’ TEE protections. (For ethical reasons, the attacks were carried out against infrastructure that was identical to but separate from the targets’ networks.) Some of the attacks included BuilderNet, dstack, and Secret Network.

BuilderNet is a network of Ethereum block builders that uses TDX to prevent parties from snooping on others’ data and to ensure fairness and that proof currency is redistributed honestly. The network builds blocks valued at millions of dollars each month.

“We demonstrated that a malicious operator with an attestation key could join BuilderNet and obtain configuration secrets, including the ability to decrypt confidential orderflow and access the Ethereum wallet for paying validators,” the TEE.fail website explained. “Additionally, a malicious operator could build arbitrary blocks or frontrun (i.e., construct a new transaction with higher fees to ensure theirs is executed first) the confidential transactions for profit while still providing deniability.”

To date, the researchers said, BuilderNet hasn’t provided mitigations. Attempts to reach BuilderNet officials were unsuccessful.

dstack is a tool for building confidential applications that run on top of virtual machines protected by Nvidia Confidential Compute. The researchers used TEE.fail to forge attestations certifying that a workload was performed by the TDX using the Nvidia protection. It also used the “borrowed” attestations to fake ownership of GPUs that a relying party trusts.

Secret Network is a platform billing itself as the “first mainnet blockchain with privacy-preserving smart contracts,” in part by encrypting on-chain data and execution with SGX. The researchers showed that TEE.fail could extract the “Concensus Seed,” the primary network-side private key encrypting confidential transactions on the Secret Network. As noted, after learning of Wiretap, the Secret Network eliminated this possibility by establishing a “curated” allowlist of known, trusted nodes allowed on the network and suspended the acceptance of new nodes. Academic or not, the ability to replicate the attack using TEE.fail shows that Wiretap wasn’t a one-off success.

A tough nut to crack

As explained earlier, the root cause of all the TEE.fail attacks is deterministic encryption, which forms the basis for protections in all three chipmakers’ TEEs. This weaker form of encryption wasn’t always used in TEEs. When Intel initially rolled out SGX, the feature was put in client CPUs, not server ones, to prevent users from building devices that could extract copyrighted content such as high-definition video.

Those early versions encrypted no more than 256MB of RAM, a small enough space to use the much stronger probabilistic form of encryption. The TEEs built into server chips, by contrast, must often encrypt terabytes of RAM. Probabilistic encryption doesn’t scale to that size without serious performance penalties. Finding a solution that accommodates this overhead won’t be easy.

One mitigation over the short term is to ensure that each 128-bit block of ciphertext has sufficient entropy. Adding random plaintext to the blocks prevents ciphertext repetition. The researchers say the entropy can be added by building a custom memory layout that inserts a 64-bit counter with a random initial value to each 64-bit block before encrypting it.

The last countermeasure the researchers proposed is adding location verification to the attestation mechanism. While insider and supply chain attacks remain a possibility inside even the most reputable cloud services, strict policies make them much less feasible. Even those mitigations, however, don’t foreclose the threat of a government agency with a valid subpoena ordering an organization to run such an attack inside their network.

In a statement, Nvidia said:

NVIDIA is aware of this research. Physical controls in addition to trust controls such as those provided by Intel TDX reduce the risk to GPUs for this style of attack, based on our discussions with the researchers. We will provide further details once the research is published.

Intel spokesman Jerry Bryant said:

Fully addressing physical attacks on memory by adding more comprehensive confidentiality, integrity and anti-replay protection results in significant trade-offs to Total Cost of Ownership. Intel continues to innovate in this area to find acceptable solutions that offer better balance between protections and TCO trade-offs.

The company has published responses here and here reiterating that physical attacks are out of scope for both TDX and SGX

AMD didn’t respond to a request for comment.

Stuck on Band-Aids

For now, TEE.fail, Wiretap, and Battering RAM remain a persistent threat that isn’t solved with the use of default implementations of the chipmakers’ secure enclaves. The most effective mitigation for the time being is for TEE users to understand the limitations and curb uses that the chipmakers say aren’t a part of the TEE threat model. Secret Network tightening requirements for operators joining the network is an example of such a mitigation.

Moore, the founder and CEO of RunZero, said that companies with big budgets can rely on custom solutions built by larger cloud services. AWS, for example, makes use of the Nitro Card, which is built using ASIC chips that accelerate processing using TEEs. Google’s proprietary answer is Titanium.

“It’s a really hard problem,” Moore said. “I’m not sure what the current state of the art is, but if you can’t afford custom hardware, the best you can do is rely on the CPU provider’s TEE, and this research shows how weak this is from the perspective of an attacker with physical access. The enclave is really a Band-Aid or hardening mechanism over a really difficult problem, and it’s both imperfect and dangerous if compromised, for all sorts of reasons.”

Dan Goodin is Senior Security Editor at Ars Technica, where he oversees coverage of malware, computer espionage, botnets, hardware hacking, encryption, and passwords. In his spare time, he enjoys gardening, cooking, and following the independent music scene. Dan is based in San Francisco. Follow him at here on Mastodon and here on Bluesky. Contact him on Signal at DanArs.82.

New physical attacks are quickly diluting secure enclave defenses from Nvidia, AMD, and Intel Read More »

AMD shores up its budget laptop CPUs by renaming more years-old silicon

AMD, Ryzen, Tech / Beth Washington / October 28, 2025

That leaves AMD with four distinct branding tiers for laptop processors: the Ryzen AI 300 series, which uses all of the company’s latest silicon and supports Windows 11’s Copilot+ features; the Ryzen 200 series for processors originally launched in mid to late 2023 as Ryzen 7040 and Ryzen 8040; Ryzen 100 for Rembrandt-R chips first launched in 2022; and then a smattering of two-digit Ryzen and Athlon brand names for Mendocino chips.

These chips are still capable of providing a decent Windows (or Linux) experience for budget PC buyers—we were big fans of the Ryzen 6000 in particular back in the fall of 2022. But the practice of giving old chips updated labels continues to feel somewhat disingenuous, and it means that users who do want AMD’s latest CPU and GPU architectures (or neural processing units, for Copilot+ PC features) will continue to pay a premium for them.

If you want to squint hard and see an upside to this for PC buyers, it’s that if you can get a good deal on a refurbished or clearance PC using Ryzen 6000, Ryzen 7035, or Ryzen 7020 chips, you’re still technically getting the latest and greatest processors that AMD is willing to sell you. The issue, as always, is that stacking more brand names on top of old processors makes it that much more difficult to make an informed buying decision.

AMD shores up its budget laptop CPUs by renaming more years-old silicon Read More »

AMD and Sony’s PS6 chipset aims to rethink the current graphics pipeline

AMD, chipset, gaming, GPU, pipeline, raytracing, Sony / Mike M. / October 10, 2025

It feels like it was just yesterday that Sony hardware architect Mark Cerny was first teasing Sony’s “PS4 successor” and its “enhanced ray-tracing capabilities” powered by new AMD chips. Now that we’re nearly five full years into the PS5 era, it’s time for Sony and AMD to start teasing the new chips that will power what Cerny calls “a future console in a few years’ time.”

In a quick nine-minute video posted Thursday, Cerny sat down with Jack Huynh, the senior VP and general manager of AMD’s Computing and Graphics Group, to talk about “Project Amethyst,” a co-engineering effort between both companies that was also teased back in July. And while that Project Amethyst hardware currently only exists in the form of a simulation, Cerny said that the “results are quite promising” for a project that’s still in the “early days.”

Mo’ ML, fewer problems?

Project Amethyst is focused on going beyond traditional rasterization techniques that don’t scale well when you try to “brute force that with raw power alone,” Huynh said in the video. Instead, the new architecture is focused on more efficient running of the kinds of machine-learning-based neural networks behind AMD’s FSR upscaling technology and Sony’s similar PSSR system.

From the same source. Two branches. One vision.

My good friend and fellow gamer @cerny and I recently reflected on our shared journey — symbolized by these two pieces of amethyst, split from the same stone.

Project Amethyst is a co-engineering effort between @PlayStation and… pic.twitter.com/De9HWV3Ub2

— Jack Huynh (@JackMHuynh) July 1, 2025

While that kind of upscaling currently helps let GPUs pump out 4K graphics in real time, Cerny said that the “nature of the GPU fights us here,” requiring calculations to be broken up into subproblems to be handled in a somewhat inefficient parallel process by the GPU’s individual compute units.

To get around this issue, Project Amethyst uses “neural arrays” that let compute units share data and process problems like a “single focused AI engine,” Cerny said. While the entire GPU won’t be connected in this manner, connecting small sets of compute units like this allows for more scalable shader engines that can “process a large chunk of the screen in one go,” Cerny said. That means Project Amethyst will let “more and more of what you see on screen… be touched or enhanced by ML,” Huynh added.

AMD and Sony’s PS6 chipset aims to rethink the current graphics pipeline Read More »

Intel’s next-generation Panther Lake laptop chips could be a return to form

AMD, core ultra, CPUs, intel, intel core ultra, panther lake, Tech / 9u50fv / October 9, 2025

Intel says that systems with these chips in them should be shipping by the end of the year. In recent years, the company has launched a small handful of ultraportable-focused CPUs at the end of the year, and then followed that up with a more fully fleshed-out midrange and high-end lineup at CES in January—we’d expect Intel to stick to that basic approach here.

Panther Lake draws near

Panther Lake tries to combine different aspects of the last-generation Lunar Lake and Arrow Lake chips. Intel

Intel’s first Core Ultra chips, codenamed Meteor Lake, were introduced two years ago. There were three big changes that separated these from the 14th-generation Core CPUs and their predecessors: They were constructed of multiple silicon tiles, fused together into one with Intel’s Foveros packaging technologies; some of those tiles were manufactured by TSMC rather than Intel; and they added a neural processing unit (NPU) that could be used for on-device machine learning and generative AI applications.

The second-generation Core Ultra chips continued to do all three of those things, but Intel pursued an odd bifurcated strategy that gave different Core Ultra 200-series processors significantly different capabilities.

The most interesting models, codenamed Lunar Lake (aka Core Ultra 200V), integrated the system RAM on the CPU package, which improved performance and power consumption while making them more expensive to buy and complicated to manufacture. These chips included Intel’s most up-to-date Arc GPU architecture, codenamed Battlemage, plus an NPU that met the performance requirements for Microsoft’s Copilot+ PC initiative.

But Core Ultra 200V chips were mostly used in high-end thin-and-light laptops. Lower-cost and higher-performance laptops got the other kind of Core Ultra 200 chip, codenamed Arrow Lake, which was a mishmash of old and new. The CPU cores used the same architecture as Lunar Lake, and there were usually more of them. But the GPU architecture was older and slower, and the NPU didn’t meet the requirements for Copilot+. If Lunar Lake was all-new, Arrow Lake was mostly an updated CPU design fused to a tweaked version of the original Meteor Lake design (confused by all these lakes yet? Welcome to my world).

Intel’s next-generation Panther Lake laptop chips could be a return to form Read More »

AMD wins massive AI chip deal from OpenAI with stock sweetener

AI, AI chips, AI development tools, AI GPU, AI infrastructure, AMD, Biz & IT, Broadcom, GPUs, machine learning, microsoft, NVIDIA, openai / Kelly Newman / October 6, 2025

As part of the arrangement, AMD will allow OpenAI to purchase up to 160 million AMD shares at 1 cent each throughout the chips deal.

OpenAI diversifies its chip supply

With demand for AI compute growing rapidly, companies like OpenAI have been looking for secondary supply lines and sources of additional computing capacity, and the AMD partnership is part the company’s wider effort to secure sufficient computing power for its AI operations. In September, Nvidia announced an investment of up to $100 billion in OpenAI that included supplying at least 10 gigawatts of Nvidia systems. OpenAI plans to deploy a gigawatt of Nvidia’s next-generation Vera Rubin chips in late 2026.

OpenAI has worked with AMD for years, according to Reuters, providing input on the design of older generations of AI chips such as the MI300X. The new agreement calls for deploying the equivalent of 6 gigawatts of computing power using AMD chips over multiple years.

Beyond working with chip suppliers, OpenAI is widely reported to be developing its own silicon for AI applications and has partnered with Broadcom, as we reported in February. A person familiar with the matter told Reuters the AMD deal does not change OpenAI’s ongoing compute plans, including its chip development effort or its partnership with Microsoft.

AMD wins massive AI chip deal from OpenAI with stock sweetener Read More »

Framework Laptop 16 update brings Nvidia GeForce to the modular gaming laptop

AMD, framework, framework laptop, framework laptop 16, gaming, NVIDIA, Ryzen, Tech / DJ Henderson / August 26, 2025

It’s been a busy year for Framework, the company behind the now well-established series of repairable, upgradeable, modular laptops (and one paradoxically less-upgradeable desktop). The company has launched a version of the Framework Laptop 13 with Ryzen AI processors, the new Framework Laptop 12, and the aforementioned desktop in the last six months, and last week, Framework teased that it still had “something big coming.”

That “something big” turns out to be the first-ever update to the Framework Laptop 16, Framework’s more powerful gaming-laptop-slash-mobile-workstation. Framework is updating the laptop with Ryzen AI processors and new integrated Radeon GPUs and is introducing a new graphics module with the mobile version of Nvidia’s GeForce RTX 5070—one that’s also fully compatible with the original Laptop 16, for upgraders.

Preorders for the new laptop open today, and pricing starts at $1,499 for a DIY Edition without RAM, storage, an OS, or Expansion Cards, a $100 increase from the price of the first Framework Laptop 16. The first units will begin shipping in November.

While Framework has launched multiple updates for its original Laptop 13, this is the first time it has updated the hardware of one of its other computers. We wouldn’t expect the just-launched Framework Laptop 12 or Framework Desktop to get an internal overhaul any time soon, but the Laptop 16 will be pushing 2-years-old by the time this upgrade launches.

The old Ryzen 7 7840HS CPU version of the Laptop 16 will still be available going forward at a slightly reduced starting price of $1,299 (for the DIY edition, before RAM and storage). The Ryzen 9 7940HS model will stick around until it sells out, at which point Framework says it’s going away.

GPU details and G-Sync asterisks

The Laptop 16’s new graphics module and cooling system, also exploded. Credit: Framework

This RTX 5070 graphics module includes a redesigned heatsink and fan system, plus an additional built-in USB-C port that supports both display output and power input (potentially freeing up one of your Expansion Card slots for something else). Because of the additional power draw of the GPU and the other new components, Framework is switching to a 240 W default power supply for the new Framework Laptop 16, up from the previous 180 W power brick.

Framework Laptop 16 update brings Nvidia GeForce to the modular gaming laptop Read More »

Ars Technica System Guide: Five sample PC builds, from $500 to $5,000

AMD, Ars System Guide, Features, gaming, intel, NVIDIA, Tech / DJ Henderson / August 18, 2025

Despite everything, it’s still possible to build decent PCs for decent prices.

You can buy a great 4K gaming PC for less than it costs to buy a GeForce RTX 5090. Let us show you some examples. Credit: Andrew Cunningham

Sometimes I go longer than I intend without writing an updated version of our PC building guide. And while I could just claim to be too busy to spend hours on Newegg or Amazon or other sites digging through dozens of near-identical parts, the lack of updates usually correlates with “times when building a desktop PC is actually a pain in the ass.”

Through most of 2025, fluctuating and inflated graphics card pricing and limited availability have once again conspired to make a normally fun hobby an annoying slog—and honestly kind of a bad way to spend your money, relative to just buying a Steam Deck or something and ignoring your desktop for a while.

But three things have brought me back for another round. First, GPU pricing and availability have improved a little since early 2025. Second, as unreasonable as pricing is for PC parts, pre-built PCs with worse specs and other design compromises are unreasonably priced, too, and people should have some sense of what their options are. And third, I just have the itch—it’s been a while since I built (or helped someone else build) a PC, and I need to get it out of my system.

So here we are! Five different suggestions for builds for a few different budgets and needs, from basic browsing to 4K gaming. And yes, there is a ridiculous “God Box,” despite the fact that the baseline ridiculousness of PC building is higher than it was a few years ago.

Notes on component selection

Part of the fun of building a PC is making it look the way you want. We’ve selected cases that will physically fit the motherboards and other parts we’re recommending and which we think will be good stylistic fits for each system. But there are many cases out there, and our picks aren’t the only options available.

It’s also worth trying to build something that’s a little future-proof—one of the advantages of the PC as a platform is the ability to swap out individual components without needing to throw out the entire system. It’s worth spending a little extra money on something you know will be supported for a while. Right this minute, that gives an advantage to AMD’s socket AM5 ecosystem over slightly cheaper but fading or dead-end platforms like AMD’s socket AM4 and Intel’s LGA 1700 or (according to rumors) LGA 1851.

As for power supplies, we’re looking for 80 Plus certified power supplies from established brands with positive user reviews on retail sites (or positive professional reviews, though these can be somewhat hard to come by for any given PSU these days). If you have a preferred brand, by all means, go with what works for you. The same goes for RAM—we’ll recommend capacities and speeds, and we’ll link to kits from brands that have worked well for us in the past, but that doesn’t mean they’re better than the many other RAM kits with equivalent specs.

For SSDs, we mostly stick to drives from known brands like Samsung, Crucial, Western Digital, and SK hynix. Our builds also include built-in Bluetooth and Wi-Fi, so you don’t need to worry about running Ethernet wires and can easily connect to Bluetooth gamepads, keyboards, mice, headsets, and other accessories.

We also haven’t priced in peripherals like webcams, monitors, keyboards, or mice, as we’re assuming most people will reuse what they already have or buy those components separately. If you’re feeling adventurous, you could even make your own DIY keyboard! If you need more guidance, Kimber Streams’ Wirecutter keyboard guides are exhaustive and educational, and Wirecutter has some monitor-buying advice, too.

Finally, we won’t be including the cost of a Windows license in our cost estimates. You can pay many different prices for Windows—$139 for an official retail license from Microsoft, $120 for an “OEM” license for system builders, or anywhere between $15 and $40 for a product key from shady gray market product key resale sites. Windows 10 keys will also work to activate Windows 11, though Microsoft stopped letting old Windows 7 and Windows 8 keys activate new Windows 10 and 11 installs a couple of years ago. You could even install Linux, given recent advancements in game compatibility layers! But if you plan to go that route, know that AMD’s graphics cards tend to be better-supported than Nvidia’s.

The budget all-rounder

What it’s good for: Browsing, schoolwork or regular work, amateur photo or video editing, and very light casual gaming. A low-cost, low-complexity introduction to PC building.

What it sucks at: You’ll need to use low settings at best for modern games, and it’s hard to keep costs down without making big sacrifices.

Cost as of this writing: $479 to $504, depending on your case

The entry point for a basic desktop PC from Dell, HP, and Lenovo is somewhere between $400 and $500 as of this writing. You can beat that pricing with a self-built one if you cut your build to the bone, and you can find tons of cheap used and refurbished stuff and serviceable mini PCs for well under that price, too. But if you’re chasing the thrill of the build, we can definitely match the big OEMs’ pricing while doing better on specs and future-proofing.

The AMD Ryzen 5 8500G should give you all the processing power you need for everyday computing and less-demanding games, despite most of its CPU cores using the lower-performing Zen 4c variant of AMD’s last-gen CPU architecture. The Radeon 740M GPU should do a decent job with many games at lower settings; it’s not a gaming GPU, but it will handle kid-friendly games like Roblox or Minecraft or undemanding battle royale or MOBA games like Fortnite and DOTA 2.

The Gigabyte B650M Gaming Plus WiFi board includes Wi-Fi, Bluetooth, and extra RAM and storage slots for future expandability. Most companies that make AM5 motherboards are pretty good about releasing new BIOS updates that patch vulnerabilities and add support for new CPUs, so you shouldn’t have a problem popping in a new processor a few years down the road if this one is no longer meeting your needs.

An AMD Ryzen 7 8700G. The 8500G is a lower-end relative of this chip, with good-enough CPU and GPU performance for light work. Credit: Andrew Cunningham

This system is spec’d for general usage and exceptionally light gaming, and 16GB of RAM and a 500 GB SSD should be plenty for that kind of thing. You can get the 1TB version of the same SSD for just $20 more, though—not a bad deal if you think light gaming is in the cards. The 600 W power supply is overkill, but it’s just $5 more than the 500 W version of the same PSU, and 600 W is enough headroom to add a GeForce RTX 4060 or 5060-series card or a Radeon RX 9600 XT to the build later on without having to worry.

The biggest challenge when looking for a decent, cheap PC case is finding one without a big, tacky acrylic window. Our standby choice for the last couple of years has been the Thermaltake Versa H17, an understated and reasonably well-reviewed option that doesn’t waste internal space on legacy features like external 3.5 and 5.25-inch drive bays or internal cages for spinning hard drives. But stock seems to be low as of this writing, suggesting it could be unavailable soon.

We looked for some alternatives that wouldn’t be a step down in quality or utility and which wouldn’t drive the system’s total price above $500. YouTubers and users generally seem to like the $70 Phanteks XT Pro, which is a lot bigger than this motherboard needs but is praised for its airflow and flexibility (it has a tempered glass side window in its cheapest configuration, and a solid “silent” variant will run you $88). The Fractal Design Focus 2 is available with both glass and solid side panels for $75.

The budget gaming PC

What it’s good for: Solid all-round performance, plus good 1080p (and sometimes 1440p) gaming performance.

What it sucks at: Future proofing, top-tier CPU performance.

Cost as of this writing: $793 to $828, depending on components

Budget gaming PCs are tough right now, but my broad advice would be the same as it’s always been: Go with the bare minimum everywhere you can so you have more money to spend on the GPU. I went into this totally unsure if I could recommend a PC I’d be happy with for the $700 to $800 we normally hit, and getting close to that number meant making some hard decisions.

I talked myself into a socket AM5 build for our non-gaming budget PC because of its future proof-ness and its decent integrated GPU, but I went with an Intel-based build for this one because we didn’t need the integrated GPU for it and because AMD still mostly uses old socket AM4 chips to cover the $150-and-below part of the market.

Given the choice between aging AMD CPUs and aging Intel CPUs, I have to give Intel the edge, thanks to the Core i5-13400F’s four E-cores. And if a 13th-gen Core chip lacks cutting-edge performance, it’s plenty fast for a midrange GPU. The $109 Core i5-12400F would also be OK and save a little more money, but we think the extra cores and small clock speed boost are worth the $20-ish premium.

For a budget build, we think your best strategy is to save money everywhere you can so you can squeeze a 16GB AMD Radeon RX 9060 XT into the budget. Credit: Andrew Cunningham

Going with a DDR4 motherboard and RAM saves us a tiny bit, and we’ve also stayed at 16GB of RAM instead of stepping up (some games, sometimes can benefit from 32GB, especially if you want to keep a bunch of other stuff running in the background, but it still usually won’t be a huge bottleneck). We upgraded to a 1TB SSD; huge AAA games will eat that up relatively quickly, but there is another M.2 slot you can use to put in another drive later. The power supply and case selections are the same as in our budget pick.

All of that cost-cutting was done in service of stretching the budget to include the 16GB version of AMD’s Radeon RX 9060 XT graphics card.

You could go with the 8GB version of the 9060 XT or Nvidia’s GeForce RTX 5060 and get solid 1080p gaming performance for almost $100 less. But we’re at a point where having 8GB of RAM in your graphics card can be a bottleneck, and that’s a problem that will only get worse over time. The 9060 XT has a consistent edge over the RTX 5060 in our testing, even in games with ray-tracing effects enabled, and at 1440p, the extra memory can easily be the difference between a game that runs and a game that doesn’t.

A more future-proofed budget gaming PC

What it’s good for: Good all-round performance with plenty of memory and storage, plus room for future upgrades.

What it sucks at: Getting you higher frame rates than our budget-budget build.

Cost as of this writing: $1,070 to $1,110, depending on components

As I found myself making cut after cut to maximize the fps-per-dollar we could get from our budget gaming PC, I decided I wanted to spec out a system with the same GPU but with other components that would make it better for non-gaming use and easier to upgrade in the future, with more generous allotments of memory and storage.

This build shifts back to many of the AMD AM5 components we used in our basic budget build, but with an 8-core Ryzen 7 7700X CPU at its heart. Its Zen 4 architecture isn’t the latest and greatest, but Zen 5 is a modest upgrade, and you’ll still get better single- and multi-core processor performance than you do with the Core i5 in our other build. It’s not worth spending more than $50 to step up to a Ryzen 7 9700X, and it’s overkill to spend $330 on a 12-core Ryzen 9 7900X or $380 on a Ryzen 7 7800X3D.

This chip doesn’t come with its own fan, so we’ve included an inexpensive air cooler we like that will give you plenty of thermal headroom.

A 32GB kit of RAM and 2TB of storage will give you ample room for games and enough RAM that you won’t have to worry about the small handful of outliers that benefit from more than 16GB of system RAM, while a marginally beefier power supply gives you a bit more headroom for future upgrades while still keeping costs relatively low.

This build won’t benefit your frame rates much since we’re sticking with the same 16GB RX 9060 XT. But the rest of it is specced generously enough that you could add a GeForce RTX 5070 (currently around $550) or a non-XT Radeon RX 9070 card (around $600) without needing to change any of the other components.

A comfortable 4K gaming rig

What it’s good for: Just about anything! But it’s built to play games at higher resolutions than our budget builds.

What it sucks at: Getting you top-of-the-line bragging rights.

Cost as of this writing: $1,829 to $1,934, depending on components.

Our budget builds cover 1080p-to-1440p gaming, and with an RTX 5070 or an RX 9070, they could realistically stretch to 4K in some games. But for more comfortable 4K gaming or super-high-frame-rate 1440p performance, you’ll thank yourself for spending a bit more.

You’ll note that the quality of the component selections here has been bumped up a bit all around. X670 or X870-series boards don’t just get you better I/O; they’ll also get you full PCI Express 5.0 support in the GPU slot and components better-suited to handling faster and more power-hungry components. We’ve swapped to a modular ATX 3.x-compliant power supply to simplify cable management and get a 12V-2×6 power connector. And we picked out a slightly higher-end SSD, too. But we’ve tried not to spend unnecessary money on things that won’t meaningfully improve performance—no 1,000+ watt power supplies, PCIe 5.0 SSDs, or 64GB RAM kits here.

A Ryzen 7 7800X3D might arguably be overkill for this build—especially at 4K, where the GPU will still be the main bottleneck—but it will be useful for getting higher frame rates at lower resolutions and just generally making sure performance stays consistent and smooth. Ryzen 7900X, 7950X, or 9900X chips are all good alternatives if you want more multi-core CPU performance—if you plan to stream as you play, for instance. A 9700X or even a 7700X would probably hold up fine if you won’t be doing that kind of thing and want to save a little.

You could cool any of these with a closed-loop AIO cooler, but a solid air cooler like the Thermalright model will keep it running cool for less money, and with a less-complicated install process.

A GeForce RTX 5070 Ti is the best 4K performance you can get for less than $1,000, but that doesn’t make it cheap. Credit: Andrew Cunningham

Based on current pricing and availability, I think the RTX 5070 Ti makes the most sense for a non-absurd 4K-capable build. Its prices are still elevated slightly above its advertised $749 MSRP, but it’s giving you RTX 4080/4080 Super-level performance for between $200 and $400 less than those cards launched for. Nvidia’s next step up, the RTX 5080, will run you at least $1,200 or $1,300—and usually more. AMD’s best option, the RX 9070 XT, is a respectable contender, and it’s probably the better choice if you plan on using Linux instead of Windows. But for a Windows-based gaming box, Nvidia still has an edge in games with ray-tracing effects enabled, plus DLSS upscaling and frame generation.

Is it silly that the GPU costs as much as our entire budget gaming PC? Of course! But it is what it is.

Even more than the budget-focused builds, the case here is a matter of personal preference, and $100 or $150 is enough to buy you any one of several dozen competent cases that will fit our chosen components. We’ve highlighted a few from case makers with good reputations to give you a place to start. Some of these also come in multiple colors, with different side panel options and both RGB and non-RGB options to suit your tastes.

If you like something a little more statement-y, the Fractal Design North ($155) and Lian Li Lancool 217 ($120) both include the wood accents that some case makers have been pushing lately. The Fractal Design case comes with both mesh and tempered glass side panel options, depending on how into RGB you are, while the Lancool case includes a whopping five case fans for keeping your system cool.

The “God Box”

What it’s good for: Anything and everything.

What it sucks at: Being affordable.

Cost as of this writing: $4,891 to $5,146

We’re avoiding Xeon and Threadripper territory here—frankly, I’ve never even tried to do a build centered on those chips and wouldn’t trust myself to make recommendations—but this system is as fast as consumer-grade hardware gets.

An Nvidia GeForce RTX 5090 guarantees the fastest GPU performance you can buy and continues the trend of “paying as much for a GPU as you could for an entire fully functional PC.” And while we have specced this build with a single GPU, the motherboard we’ve chosen has a second full-speed PCIe 5.0 x16 slot that you could use for a dual-GPU build.

A Ryzen 9950X3D chip gets you top-tier gaming performance and tons of CPU cores. We’re cooling this powerful chip with a 360 mm Arctic Liquid Freezer III Pro cooler, which has generally earned good reviews from Gamers Nexus and other outlets for its value, cooling performance, and quiet performance. A white option is also available if you’re going for a light-mode color scheme instead of our predominantly dark-mode build.

Other components have been pumped up similarly gratuitously. A 1,000 W power supply is the minimum for an RTX 5090, but to give us some headroom, why not use a 1,200 W model with lights on it? Is PCIe 5.0 storage strictly necessary for anything? No! But let’s grab a 4 TB PCIe 5.0 SSD anyway. And populating all four of our RAM slots with a 32GB stick of DDR5 avoids any unsightly blank spots inside our case.

We’ve selected a couple of largish case options to house our big builds, though as usual, there are tons of other options to fit all design sensibilities and tastes. Just make sure, if you’re selecting a big Extended ATX motherboard like the X870E Taichi, that your case will fit a board that’s slightly wider than a regular ATX or micro ATX board (the Taichi is 267 mm wide, which should be fine in either of our case selections).

Ars Technica System Guide: Five sample PC builds, from $500 to $5,000 Read More »

Trump strikes “wild” deal making US firms pay 15% tax on China chip sales

AI, AI chips, AMD, Donald Trump, Huawei, Micron, NVIDIA, Policy, semiconductor, tariffs, us-china trade war / 9u50fv / August 11, 2025

“Extra penalty” for US firms

The deal won’t resolve national security concerns.

Ahead of an August 12 deadline for a US-China trade deal, Donald Trump’s tactics continue to confuse those trying to assess the country’s national security priorities regarding its biggest geopolitical rival.

For months, Trump has kicked the can down the road regarding a TikTok ban, allowing the app to continue operating despite supposedly urgent national security concerns that China may be using the app to spy on Americans. And now, in the latest baffling move, a US official announced Monday that Trump got Nvidia and AMD to agree to “give the US government 15 percent of revenue from sales to China of advanced computer chips,” Reuters reported. Those chips, about 20 policymakers and national security experts recently warned Trump, could be used to fuel China’s frontier AI, which seemingly poses an even greater national security risk.

Trump’s “wild” deal with US chip firms

Reuters granted two officials anonymity to discuss Trump’s deal with US chipmakers, because details have yet to be made public. Requiring US firms to pay for sales in China is an “unusual” move for a president, Reuters noted, and the Trump administration has yet to say what exactly it plans to do with the money.

For US firms, the deal may set an alarming precedent. Not only have analysts warned that the deal could “hurt margins” for both companies, but export curbs on Nvidia’s H20 chips, for example, had been established to prevent US technology thefts, secure US technology leadership, and protect US national security. Now the US government appears to be accepting a payment to overlook those alleged risks, without much reassurance that the policy won’t advantage China in the AI race.

The move drew immediate scrutiny from critics, including Geoff Gertz, a senior fellow at the US think tank Center for a New American Security, who told Reuters that he thinks the deal is “wild.”

“Either selling H20 chips to China is a national security risk, in which case we shouldn’t be doing it to begin with, or it’s not a national security risk, in which case, why are we putting this extra penalty on the sale?” Gertz posited.

At this point, the only reassurance from the Trump administration is an official suggesting (without providing any rationale) that selling H20 or equivalent chips—which are not Nvidia’s most advanced chips—no longer compromises national security.

Trump “trading away” national security

It remains unclear when or how the levy will be implemented.

For chipmakers, the levy is likely viewed as a relatively small price to pay to avoid export curbs. Nvidia had forecasted $8 billion in potential losses if it couldn’t sell its H20 chips to China. AMD expected $1 billion in revenue cuts, partly due to the loss of sales for its MI308 chips in China.

The firms apparently agreed to Trump’s deal as a condition to receive licenses to export those chips. But caving to Trump could bite them back in the long run, AJ Bell, investment director Russ Mould, told Reuters—perhaps especially if Trump faces increasing pressure over feared national security concerns.

“The Chinese market is significant for both these companies, so even if they have to give up a bit of the money, they would otherwise make it look like a logical move on paper,” Mould said. However, the deal “is unprecedented and there is always the risk the revenue take could be upped or that the Trump administration changes its mind and re-imposes export controls.”

So far, AMD has not commented on the report. Nvidia’s spokesperson declined to comment beyond noting, “We follow rules the US government sets for our participation in worldwide markets.”

A former adviser to Joe Biden’s Commerce Department, Alasdair Phillips-Robins, told Reuters that the levy suggests the Trump administration “is trading away national security protections for revenue for the Treasury.”

Huawei close to unveiling new AI chip tech

The end of a 90-day truce between the US and China is rapidly approaching, with the US signaling that the truce will likely be extended soon as Trump attempts to get a long-sought-after meeting with China’s President Xi Jinping.

For China, gutting export curbs on chips remains a key priority in negotiations, the Financial Times reported Sunday. But Nvidia’s H20 chips, for example, are lower priority than high-bandwidth memory (HBM) chips, sources told FT.

Chinese state media has even begun attacking the H20 chips as a Chinese national security risk. It appears that China is urging a boycott on H20 chips due to questions linked to a recent Congressional push to require chipmakers to build “backdoors” that would allow remote shutdowns of any chips detected as non-compliant with export curbs. That bill may mean that Nvidia’s chips already allow for US surveillance, China seemingly fears. (Nvidia has denied building such backdoors.)

Biden banned HBM exports to China last year, specifically moving to hamper innovation of Chinese chipmakers Huawei and Semiconductor Manufacturing International Corporation (SMIC).

Currently, US firms AMD and Micron remain top suppliers of HBM chips globally, along with South Korean firms Samsung Electronics and SK Hynix, but Chinese firms have notably lagged behind, South China Morning Post (SCMP) reported. One source told FT that China “had raised the HBM issue in some” Trump negotiations, likely directly seeking to lift Biden’s “HBM controls because they seriously constrain the ability of Chinese companies, including Huawei, to develop their own AI chips.”

For Trump, the HBM controls could be seen as leverage to secure another trade win. However, some experts are hoping that Trump won’t play that card, citing concerns from the Biden era that remain unaddressed.

If Trump bends to Chinese pressure and lifts HBM controls, China could more easily produce AI chips at scale, Biden had feared. That could even possibly endanger US firms’ standing as world leaders, seemingly including threatening Nvidia, a company that Trump discovered this term. Gregory Allen, an AI expert at a US think tank called the Center for Strategic and International Studies, told FT that “saying that we should allow more advanced HBM sales to China is the exact same as saying that we should help Huawei make better AI chips so that they can replace Nvidia.”

Meanwhile, Huawei is reportedly already innovating to help reduce China’s reliance on HBM chips, the SCMP reported on Monday. Chinese state-run Securities Times reported that Huawei is “set to unveil a technological breakthrough that could reduce China’s reliance on high-bandwidth memory (HBM) chips for running artificial intelligence reasoning models” at the 2025 Financial AI Reasoning Application Landing and Development Forum in Shanghai on Tuesday.

It’s a conveniently timed announcement, given the US-China trade deal deadline lands the same day. But the risk of Huawei possibly relying on US tech to reach that particular milestone is why HBM controls should remain off the table during Trump’s negotiations, one official told FT.

“Relaxing these controls would be a gift to Huawei and SMIC and could open the floodgates for China to start making millions of AI chips per year, while also diverting scarce HBM from chips sold in the US,” the official said.

Experts and policymakers had previously warned Trump that allowing H20 export curbs could similarly reduce access to semiconductors in the US, potentially disrupting the entire purpose of Trump’s trade war, which is building reliable US supply chains. Additionally, allowing exports will likely drive up costs to US chip firms at a time when they noted “projected data center demand from the US power market would require 90 percent of global chip supply through 2030, an unlikely scenario even without China joining the rush to buy advanced AI chips.” They’re now joined by others urging Trump to revive Biden’s efforts to block chip exports to China, or else risk empowering a geopolitical rival to become a global AI leader ahead of the US.

Ashley is a senior policy reporter for Ars Technica, dedicated to tracking social impacts of emerging policies and new technologies. She is a Chicago-based journalist with 20 years of experience.

Trump strikes “wild” deal making US firms pay 15% tax on China chip sales Read More »

Review: Framework Desktop is a mash-up of a regular desktop PC and the Mac Studio

AMD, framework, framework desktop, gaming, NVIDIA, ryzen ai, Tech / DJ Henderson / August 7, 2025

Size matters most for Framework’s first stab at a desktop workstation/gaming PC.

The Framework Desktop. Credit: Andrew Cunningham

Framework’s main claim to fame is its commitment to modular, upgradeable, repairable laptops. The jury’s still out on early 2024’s Framework Laptop 16 and mid-2025’s Framework Laptop 12, neither of which has seen a hardware refresh, but so far, the company has released half a dozen iterations of its flagship Framework Laptop 13 in less than five years. If you bought one of the originals right when it first launched, you could go to Framework’s site, buy an all-new motherboard and RAM, and get a substantial upgrade in performance and other capabilities without having to change anything else about your laptop.

Framework’s laptops haven’t been adopted as industry-wide standards, but in many ways, they seem built to reflect the flexibility and modularity that has drawn me to desktop PCs for more than two decades.

That’s what makes the Framework Desktop so weird. Not only is Framework navigating into a product category where its main innovation and claim to fame is totally unnecessary. But it’s actually doing that with a desktop that’s less upgradeable and modular than any given self-built desktop PC.

The Framework Desktop has a lot of interesting design touches, and it’s automatically a better buy than the weird AMD Ryzen AI Max-based mini desktops you can buy from a couple of no-name manufacturers. But aside from being more considerate of PC industry standards, the Framework Desktop asks the same question that any gaming-focused mini PC does: Do you care about having a small machine so much that you would pay more money for less performance, and for a system you can’t upgrade much after you buy it?

Design and assembly

Opening the Framework Desktop’s box. The PC and all its accessories are neatly packed away in all-recyclable carboard and paper. Andrew Cunningham

My DIY Edition Framework Desktop arrived in a cardboard box that was already as small or a bit smaller than my usual desktop PC, a mini ITX build with a dedicated GPU inside a 14.67-liter SSUPD Meshlicious case. It’s not a huge system, especially for something that can fit a GeForce RTX 5090 in it. But three of the 4.5-liter Framework Desktops could fit inside my build’s case with a little space leftover.

The PC itself is buried a couple of layers deep in this box under some side panels and whatever fan you choose (Framework offers RGB and non-RGB options from Cooler Master and Noctua, but any 120 mm fan will fit on the heatsink). Even for the DIY Edition, the bulk of it is already assembled: the motherboard is in the case, a large black heatsink is already perched atop the SoC, and both the power supply and front I/O ports are already hooked up.

The aspiring DIYer mainly needs to install the SSD and the fan to get going. Putting in these components gives you a decent crash course in how the system goes together and comes apart. The primary M.2 SSD slot is under a small metal heat spreader next to the main heatsink—loosen one screw to remove it, and install your SSD of choice. The system’s other side panel can be removed to expose a second M.2 SSD slot and the Wi-Fi/Bluetooth module, letting you install or replace either.

Lift the small handles on the two top screws and loosen them by hand to remove them, and the case’s top panel slides off. This provides easier access to both the CPU fan header and RGB header, so you can connect the fan after you install it and its plastic shroud on top of the heatsink. That’s pretty much it for assembly, aside from sliding the various panels back in place to close the thing up and reinstalling the top screws (or, if you bought or printed one, adding a handle to the top of the case).

The Framework Desktop includes a beefier version of Framework’s usual screwdriver with a longer bit. Credit: Andrew Cunningham

Framework includes a beefier version of its typical screwdriver with the Desktop, including a bit that can be pulled out and reversed to be switched between Phillips and Torx heads. The iFixit-style install instructions are clearly written and include plenty of high-resolution sample images so you can always tell how things are supposedto look.

The front of the system requires some assembly, too, but all of this stuff can be removed and replaced easily without opening up the rest of the system. The front panel, where the system’s customizable tiles can be snapped on and popped off, attaches with magnets and can easily be pried away from the desktop with your fingernails. At the bottom are slots for two of Framework’s USB-C Expansion Cards, the same ones that all the Framework Laptops use.

By default, those ports are limited to 5 Gbps USB transfer speeds in the BIOS, something the system says reduces wireless interference; those with all-wired networking and accessories can presumably enable the full 10 Gbps speeds without downsides. The front ports should support all of the Expansion Cards except for display outputs, which they aren’t wired for. (I also had issues getting the Desktop to boot from a USB port on the front of the system while installing Windows, but your mileage may vary; using one of the rear USB ports solved the issue for me.)

Standards, sometimes

Putting in the M.2 SSD. There’s another SSD slot on the back of the motherboard. Andrew Cunningham

What puts the Framework Desktop above mini PCs from Amazon or the various gaming NUCs that Intel and Asus have released over the years is a commitment to standards.

For reasons we’ll explore later, there was no way to build the system around this specific AMD chip without using soldered-on memory. But the motherboard is a regular mini ITX-sized motherboard. Other ITX boards will fit into Framework’s case, and the Framework Laptop’s motherboard will fit into other systems (as long as they can also fit the fan and heatsink).

The 400 W power supply conforms to the FlexATX standard. The CPU fan is just a regular 120 mm fan, and the mounting holes for system fans on the front can take any 92 mm fan. The two case fan headers on the motherboard are the same ones you’d find on any motherboard you bought for yourself. The front panel ports can’t be used for display outputs, but anything else ought to work.

Few elements of the Framework Desktop are truly proprietary, and if Framework went out of business tomorrow, you’d still have a lot of flexibility for buying and installing replacement parts. The problem is that the soldered-down, non-replaceable, non-upgradeable parts are the CPU, GPU, and RAM. There’s at least a little flexibility with the graphics card if you move the board into a different case—there’s a single PCIe x4 slot on the board that you could put an external GPU into, though many PCIe x16 graphics cards will be bandwidth starved. But left in its original case, it’s an easy-to-work-on, standards-compliant system that will also never be any better or get any faster than it is the day you buy it.

Hope you like plastic

Snapping some tiles into the Framework Desktop’s plastic front panel. Credit: Andrew Cunningham

The interior of the Framework Desktop is built of sturdy metal, thoughtfully molded to give easy access to each of the ports and components on the motherboard. My main beef with the system is the outside.

The front and side panels of the Framework Desktop are all made out of plastic. The clear side panel, if you spring for it, is made of a thick acrylic instead of tempered glass (presumably because Framework has drilled holes in the side of it to improve airflow).

This isn’t the end of the world, but the kinds of premium ITX PC cases that the Desktop is competing with are predominantly made of nicer-looking and nicer-feeling metal rather than plastic. It just feels surprisingly cheap, which was an unpleasant surprise—even the plastic Framework Laptop 12 felt sturdy and high-quality, something I can’t really say of the Desktop’s exterior panels.

I do like the design on the front panel—a grid of 21 small square plastic tiles that users can rearrange however they want. Framework sells tiles with straight and diagonal lines on them, plus individual tiles with different logos or designs printed or embossed on them. If you install a fan in the front of the system, you’ll want to stick to the lined tiles in the top 9 x 9 section of the grid, which will allow air to pass through. The tiles with images on them are solid—putting a couple of them in front of a fan likely won’t hurt your airflow too much, but you won’t want to use too many.

Framework has also published basic templates for both the tiles and the top panel so that those with 3D printers can make their own.

PC testbed notes

We’ve compared the performance of the Framework Desktop to a bunch of other PCs to give you a sense of how it stacks up to full-size desktops. We’ve also compared it to the Ryzen 7 8700G in a Gigabyte B650I Aorus Ultra mini ITX motherboard with 32GB of DDR5-6400 to show the best performance you can expect from a similarly sized socketed desktop system.

Where possible, we’ve also included some numbers from the M4 Pro Mac mini and the M4 Max Mac Studio, two compact desktops in the same general price range as the Framework Desktop.

For our game benchmarks, the dedicated GPU results were gathered using our GPU testbed, which you can read about in our latest dedicated GPU review. The integrated GPUs were obviously tested with the CPUs they’re attached to.

	AMD AM5	Intel LGA 1851	Intel LGA 1700
CPUs	Ryzen 7000 and 9000 series	Core Ultra 200 series	12th, 13th, and 14th-generation Core
Motherboard	ASRock X870E Taichi or MSI MPG X870E Carbon Wifi (provided by AMD)	MSI MEG Z890 Unify-X (provided by Intel)	Gigabyte Z790 Aorus Master X (provided by Intel)
RAM config	32GB G.Skill Trident Z5 Neo (provided by AMD), running at DDR5-6000	32GB G.Skill Trident Z5 Neo (provided by AMD), running at DDR5-6000	32GB G.Skill Trident Z5 Neo (provided by AMD), running at DDR5-6000

Performance and power

Our Framework-provided review unit was the highest-end option; it has a 16-core Ryzen AI Max+395 processor, 40 graphics cores, and 128GB of RAM. At $1,999 before adding an SSD, a fan, an OS, front tiles, or Expansion Cards, this is the best, priciest configuration Framework offers. The $1,599 configuration uses the same chip with the same performance, but with 64GB of RAM instead.

All 16 of those CPU cores are based on the Zen 5 architecture, with none of the smaller-but-slower Zen 5c cores. But its total TDP is also limited to 120 W in total, which will hold it back a bit compared to socketed 16-core desktop CPUs like the Ryzen 9 9950X, which has a 170 W default TDP for the CPU alone.

In our testing, it seems clear that the CPU throttles when being tasked with intensive multi-core work like our Handbrake test, with temperatures that spike to around 100 degrees Celsius and hang out at around or just under that number for the duration of our test runs. The CPU package uses right around 100 W on average (this will vary based on the tests you’re running and how long you’re running them), compared to the 160 W and 194 W that the 12- and 16-core Ryzen 9 9900X and 9950X can consume at their default power levels.

Those are socketed desktop chips in huge cases being cooled by large AIO watercooling loops, so it’s hardly a fair comparison. The Framework Desktop’s CPU is also quite efficient, using even less power to accomplish our video encoding test than the 9950X in its 105 W Eco Mode. But this is the consequence of prioritizing a small size—a 16-core processor that, under heavy loads, performs more like a 12-core or even an 8-core desktop processor.

The upside is that the Framework Desktop is quieter than most desktops either under load or when idling. By default, the main CPU fan will turn off entirely when the system is under light load, and I often noticed it parking itself when I was just browsing or moving files around.

Based on our gaming tests, the Framework Desktop should be a competent 1080p-to-1440p midrange gaming system. We observed similar performance from the Radeon 8060S integrated GPU when we tested it in the Asus ROG Flow Z13 tablet. For an integrated GPU, it’s head and shoulders over anything you can get in a socketed desktop system, and it easily ran three or four times faster than the Radeon 780M in the 8700G. The soldered RAM is annoying, but the extra speed it enables helps address the memory bandwidth problem that starves most integrated GPUs.

Compared to other desktop GPUs, though, the 8060S is merely fine. It’s usually a little slower than the last-generation Radeon RX 7600 XT, a card that cost $329 when it launched in early 2024—and with a performance hit that’s slightly more pronounced in games with ray-tracing effects on.

The 8060S stacks up OK to older midrange GPUs like the GeForce RTX 3060 and 4060, but it’s soundly beaten by the RTX 5060 or the 16GB version of the Radeon RX 9060 XT, cards currently available for $300-to-$400. (One problem for the 8060S—it’s based on the RDNA3.5 architecture, so it’s missing ray-tracing performance improvements introduced in RDNA4 and the RX 9000 series).

All of that said, the GPU may be more interesting than it looks on paper for people whose workloads need gobs and gobs of graphics memory but who don’t necessarily need that memory to be attached to the blazing-fastest GPU that exists. For people running certain AI or machine learning workloads, the 8060S’s unified memory setup means you can get a GPU with 64GB or 128GB of VRAM for less than the price of a single RTX 5090 (Framework says the GPU can use up to 112GB of RAM on the 128GB Desktop). Framework is advertising that use case pretty extensively, and it offers a guide to setting up large language models to run locally on the system.

That memory would likely be even more useful if it were attached to an Nvidia GPU instead of an AMD model—Nvidia’s hold on the workstation graphics market is at least as tight as its hold on the gaming GPU market, and many apps and tools support Nvidia GPUs and CUDA first/best/only. But it’s still one possible benefit the Framework Desktop might offer, relative to a desktop with a dedicated GPU.

You can’t say it isn’t unique

The Framework Desktop is a bit like a PC tower blended with Apple’s Mac Studio. Credit: Andrew Cunningham

In one way, Framework has done the same thing with the Desktop that it has done with all its laptops: found a niche and built a product to fill it. And with its standard-size components and standard connectors, the Framework Desktop is a clear cut above every Intel gaming NUC or Asus ROG thingamajig that’s ever existed.

I’m always impressed by the creativity, thoughtfulness, and attention to detail that Framework brings to its builds. For the Desktop, this is partially offset by how much I don’t care for most of its cheap plastic-and-acrylic exterior. But it’s still thoughtfully designed on the inside, with as much respect for standards, modularity, and repairability as you can get, once you get past that whole thing where that the major functional components are all irrevocably soldered together.

The Framework Desktop is also quiet, cute, and reasonably powerful. You’re paying some extra money and giving up both CPU and GPU speed to get something small. But you won’t run into games or apps that simply refuse to run for performance-related reasons.

It does feel like a weird product for Framework to build, though. It’s not that I can’t imagine the kind of person a Framework Desktop might be good for—it’s that I think Framework has built its business targeting a PC enthusiast demographic that will mainly be turned off by the desktop’s lack of upgradeability.

The Framework desktop is an interesting option for people who want or need a compact and easy-to-build workstation or gaming PC, or a Windows-or-Linux version of Apple’s Mac Studio. It will fit comfortably under a TV or in a cramped office. It’s too bad that it isn’t easier to upgrade. But for people who would prefer the benefits of a socketed CPU or a swappable graphics card, I’m sure the people at Framework would be the first ones to point you in the direction of a good-old desktop PC.

The good

Solid all-round performance and good power efficiency.
The Radeon 8060S is exceptionally good for an integrated GPU, delivering much better performance than you can get in something like the Ryzen 7 8700G.
Large pool of RAM available to the GPU could be good for machine learning and AI workloads.
Thoughtfully designed interior that’s easy to put together.
Uses standard-shaped motherboard, fan headers, power supply, and connectors, unlike lots of pre-built mini PCs.
Front tiles are fun.

The bad

Power limits keep the 16-core CPU from running as fast as the socketed desktop version.
A $300-to-$400 dedicated GPU will still beat the Radeon RX 8060S.
Cheap-looking exterior plastic panels.

The ugly

Soldered RAM in a desktop system.

Review: Framework Desktop is a mash-up of a regular desktop PC and the Mac Studio Read More »

AMD’s $299 Radeon RX 9060 XT brings 8GB or 16GB of RAM to fight the RTX 5060

AMD, gaming, Radeon, RNDA 4, Tech / Mike M. / May 21, 2025

AMD didn’t provide much by way of performance comparisons, but it’s promising that the cards have the same number of compute units as AMD’s last-generation RX 7600 series. AMD says that RDNA 4 compute units are much faster than those used for RDNA 3, particularly in games with ray-tracing effects enabled. This helped make the Radeon RX 9070 cards generally as fast or faster than the RX 7900 XTX and 7900 XT series, despite having around two-thirds as many compute units. Sticking with 32 CUs for the 9060 series isn’t exciting on paper, but we should still see a respectable generation-over-generation performance bump. The RX 7600 series, by contrast, provided a pretty modest performance improvement compared to 2022’s Radeon RX 6650 XT.

AMD says that the cards’ total board power—the amount of power the entire graphics card, including the GPU itself, RAM, and other components—starts at 150 W for the 8GB card and 160 W for the 16GB card, with a maximum TBP of 182 W. That’s a shade higher than but generally comparable to the RTX 5060 and 5060 Ti, and (depending on where actual performance ends up) quite a bit more efficient than the RX 7600 series. This partly comes down to a more efficient 4nm TSMC manufacturing process, a substantial upgrade from the 6nm process used for the 7600 series.

It’s unusual for a GPU maker to define a TBP range—more commonly we’re just given a single default value. But this is in line with new settings we observed in our RX 9070 review; AMD officially supports a range of different user-selectable TBP numbers in its Catalyst driver package, and some GPU makers were shipping cards that used higher TBPs by default.

Higher power limits can increase performance, though usually the performance increase is disproportionately small compared to the increase in power draw. These power limits should also generally mean that most 9060 XTs can be powered with a single 8-pin power connector, rather than using multiple connectors or the 12-pin 12VHPWR/12V-2×6 connector.

AMD’s $299 Radeon RX 9060 XT brings 8GB or 16GB of RAM to fight the RTX 5060 Read More »