NVIDIA – Page 2

China tells Alibaba, ByteDance to justify purchases of Nvidia AI chips

AI, china, NVIDIA, Policy, syndication / 9u50fv / August 12, 2025

Beijing is demanding tech companies including Alibaba and ByteDance justify their orders of Nvidia’s H20 artificial intelligence chips, complicating the US chipmaker’s business in China after striking an export arrangement with the Trump administration.

The tech companies have been asked by regulators such as the Ministry of Industry and Information Technology (MIIT) to explain why they need to order Nvidia’s H20 chips instead of using domestic alternatives, said three people familiar with the situation.

Some tech companies, who were the main buyers of Nvidia’s H20 chips before their sale in China was restricted, were planning to downsize their orders as a result of the questions from regulators, said two of the people.

“It’s not banned but has kind of become a politically incorrect thing to do,” said one Chinese data center operator about purchasing Nvidia’s H20 chips.

Alibaba, ByteDance, and MIIT did not immediately respond to a request for comment.

Chinese regulators have expressed growing disapproval of companies using Nvidia’s chips for any government or security related projects. Bloomberg reported on Tuesday that Chinese authorities had sent notices to a range of companies discouraging the use of the H20 chips, particularly for government-related work.

China tells Alibaba, ByteDance to justify purchases of Nvidia AI chips 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.

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

Nvidia announces end of GPU driver updates for GeForce 10-series, Windows 10

gaming, GeForce, geforce gtx, geforce gtx 1060, NVIDIA, Tech / Shannon Garcia / July 31, 2025

The Maxwell, Pascal, and Volta GPUs won’t be totally abandoned after 2025; Nvidia says it will release quarterly security updates for these cards through October 2028. These updates won’t optimize performance or fix bugs in any new games, but if you still have an older or hand-me-down PC using one of these cards to play Minecraft or Roblox, you won’t be leaving yourself open to GPU-related security exploits.

Nvidia has dropped hints that the end of support for these older GPUs was coming. The company announced back in January that CUDA support for the Maxwell, Pascal, and Volta architectures was considered “feature complete” and was being frozen. This is the first time since 2021 that Nvidia has dropped support for older GPUs.

As for Windows 10, Microsoft has been pushing users toward Windows 11 for years, including by using full-screen ads encouraging people to buy new Copilot+ PCs, but the older operating system still has a sizable user base. According to the Steam Hardware Survey, Windows 10 is in decline, but it still powers over a third of the PCs in the survey as of June 2025, compared to a little over 60 percent for Windows 11.

Nvidia announces end of GPU driver updates for GeForce 10-series, Windows 10 Read More »

China claims Nvidia built backdoor into H20 chip designed for Chinese market

china, NVIDIA, Policy, syndication, Tech / Shannon Garcia / July 31, 2025

The CAC did not specify which experts had found a back door in Nvidia’s products or whether any tests in China had uncovered the same results. Nvidia did not immediately respond to a request for comment.

Lawmakers in Washington have expressed concern about chip smuggling and introduced a bill that would require chipmakers such as Nvidia to embed location tracking into export-controlled hardware.

Beijing has issued informal guidance to major Chinese tech groups to increase purchases of domestic AI chips in order to reduce reliance on Nvidia and support the evolution of a rival domestic chip ecosystem.

Chinese tech giant Huawei and smaller groups including Biren and Cambricon have benefited from the push to localize chip supply chains.

Nvidia said it would take nine months from restarting manufacturing to shipping the H20 to clients. Industry insiders said there was considerable uncertainty among Chinese customers over whether they would be able to take delivery of any orders if the US reversed its decision to allow its sale.

The Trump administration has faced heavy criticism, including from security experts and former officials, who argue that the H20 sales would accelerate Chinese AI development and threaten US national security.

“There are strong factions on both sides of the Pacific that don’t like the idea of renewing H20 sales,” said Triolo. “In the US, the opposition is clear, but also in China voices are saying that it will slow transition to the alternative ecosystem.”

China claims Nvidia built backdoor into H20 chip designed for Chinese market Read More »

Trump caving on Nvidia H20 export curbs may disrupt his bigger trade war

AI, AI chips, Donald Trump, frontier ai, h20 chips, Howard Lutnick, Jensen Huang, NVIDIA, Policy, semiconductors, tariffs, trump tariffs, us-china trade war / 9u50fv / July 29, 2025

But experts seem to fear that Trump isn’t paying enough attention to how exports of US technology could threaten to not only supercharge China’s military and AI capabilities but also drain supplies that US firms need to keep the US at the forefront of AI innovation.

“More chips for China means fewer chips for the US,” experts said, noting that “China’s biggest tech firms, including Tencent, ByteDance, and Alibaba,” have spent $16 billion on bulk-ordered H20 chips over the past year.

Meanwhile, “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,” experts said. If Trump doesn’t intervene, one of America’s biggest AI rivals could even end up driving up costs of AI chips for US firms, they warned.

“We urge you to reverse course,” the letter concluded. “This is not a question of trade. It is a question of national security.”

Trump says he never heard of Nvidia before

Perhaps the bigger problem for Trump, national security experts suggest, would be if China or other trade partners perceive the US resolve to wield export controls as a foreign policy tool to be “weakened” by Trump reversing course on H20 controls.

They suggested that Trump caving on H20 controls could even “embolden China to seek additional access concessions” at a time when some analysts suggest that China may already have an upper hand in trade negotiations.

The US and China are largely expected to extend a 90-day truce following recent talks in Stockholm, Reuters reported. Anonymous sources told the South China Morning Post that the US may have already agreed to not impose any new tariffs or otherwise ratchet up the trade war during that truce, but that remains unconfirmed, as Trump continues to warn that chip tariffs are coming soon.

Trump has recently claimed that he thinks he may be close to cementing a deal with China, but it appears likely that talks will continue well into the fall. A meeting between Trump and Chinese President Xi Jinping probably won’t be scheduled until late October or early November, Reuters reported.

Trump caving on Nvidia H20 export curbs may disrupt his bigger trade war Read More »

OpenAI and partners are building a massive AI data center in Texas

AI, AI chips, AI GPU, AI infrastructure, Biz & IT, datacenters, Larry Ellison, machine learning, microsoft, NVIDIA, openai, oracle, sam altman, SoftBank, Stargate / 9u50fv / July 24, 2025

Stargate moves forward despite early skepticism

When OpenAI announced Stargate in January, critics questioned whether the company could deliver on its ambitious $500 billion funding promise. Trump ally and frequent Altman foe Elon Musk wrote on X that “They don’t actually have the money,” claiming that “SoftBank has well under $10B secured.”

Tech writer and frequent OpenAI critic Ed Zitron raised concerns about OpenAI’s financial position, noting the company’s $5 billion in losses in 2024. “This company loses $5bn+ a year! So what, they raise $19bn for Stargate, then what, another $10bn just to be able to survive?” Zitron wrote on Bluesky at the time.

Six months later, OpenAI’s Abilene data center has moved from construction to partial operation. Oracle began delivering Nvidia GB200 racks to the facility last month, and OpenAI reports it has started running early training and inference workloads to support what it calls “next-generation frontier research.”

Despite the White House announcement with President Trump in January, the Stargate concept dates back to March 2024, when Microsoft and OpenAI partnered on a $100 billion supercomputer as part of a five-phase plan. Over time, the plan evolved into its current form as a partnership with Oracle, SoftBank, and CoreWeave.

“Stargate is an ambitious undertaking designed to meet the historic opportunity in front of us,” writes OpenAI in the press release announcing the latest deal. “That opportunity is now coming to life through strong support from partners, governments, and investors worldwide—including important leadership from the White House, which has recognized the critical role AI infrastructure will play in driving innovation, economic growth, and national competitiveness.”

OpenAI and partners are building a massive AI data center in Texas Read More »

Nvidia chips become the first GPUs to fall to Rowhammer bit-flip attacks

bitfl, bitflips, Biz & IT, GPUs, NVIDIA, rowhammer, Security / DJ Henderson / July 15, 2025

GPUhammer is the first to flip bits in onboard GPU memory. It likely won’t be the last.

The Nvidia RTX-A6000. Credit: Nvidia

Nvidia is recommending a mitigation for customers of one of its GPU product lines that will degrade performance by up to 10 percent in a bid to protect users from exploits that could let hackers sabotage work projects and possibly cause other compromises.

The move comes in response to an attack a team of academic researchers demonstrated against Nvidia’s RTX A6000, a widely used GPU for high-performance computing that’s available from many cloud services. A vulnerability the researchers discovered opens the GPU to Rowhammer, a class of attack that exploits physical weakness in DRAM chip modules that store data.

Rowhammer allows hackers to change or corrupt data stored in memory by rapidly and repeatedly accessing—or hammering—a physical row of memory cells. By repeatedly hammering carefully chosen rows, the attack induces bit flips in nearby rows, meaning a digital zero is converted to a one or vice versa. Until now, Rowhammer attacks have been demonstrated only against memory chips for CPUs, used for general computing tasks.

Like catastrophic brain damage

That changed last week as researchers unveiled GPUhammer, the first known successful Rowhammer attack on a discrete GPU. Traditionally, GPUs were used for rendering graphics and cracking passwords. In recent years, GPUs have become the workhorses for tasks such as high-performance computing, machine learning, neural networking, and other AI uses. No company has benefited more from the AI and HPC boom than Nvidia, which last week became the first company to reach a $4 trillion valuation. While the researchers demonstrated their attack against only the A6000, it likely works against other GPUs from Nvidia, the researchers said.

The researchers’ proof-of-concept exploit was able to tamper with deep neural network models used in machine learning for things like autonomous driving, healthcare applications, and medical imaging for analyzing MRI scans. GPUHammer flips a single bit in the exponent of a model weight—for example in y, where a floating point is represented as x times 2^y. The single bit flip can increase the exponent value by 16. The result is an altering of the model weight by a whopping 2¹⁶, degrading model accuracy from 80 percent to 0.1 percent, said Gururaj Saileshwar, an assistant professor at the University of Toronto and co-author of an academic paper demonstrating the attack.

“This is like inducing catastrophic brain damage in the model: with just one bit flip, accuracy can crash from 80% to 0.1%, rendering it useless,” Saileshwar wrote in an email. “With such accuracy degradation, a self-driving car may misclassify stop signs (reading a stop sign as a speed limit 50 mph sign), or stop recognizing pedestrians. A healthcare model might misdiagnose patients. A security classifier may fail to detect malware.”

In response, Nvidia is recommending users implement a defense that could degrade overall performance by as much as 10 percent. Among machine learning inference workloads the researchers studied, the slowdown affects the “3D U-Net ML Model” the most. This model is used for an array of HPC tasks, such as medical imaging.

The performance hit is caused by the resulting reduction in bandwidth between the GPU and the memory module, which the researchers estimated as 12 percent. There’s also a 6.25 percent loss in memory capacity across the board, regardless of the workload. Performance degradation will be the highest for applications that access large amounts of memory.

A figure in the researchers’ academic paper provides the overhead breakdowns for the workloads tested.

Overheads of enabling ECC in A6000 GPU for MLPerf Inference and CUDA samples benchmarks. Credit: Lin et al.

Rowhammer attacks present a threat to memory inside the typical laptop or desktop computer in a home or office, but most Rowhammer research in recent years has focused on the threat inside cloud environments. That’s because these environments often allot the same physical CPU or GPU to multiple users. A malicious attacker can run Rowhammer code on a cloud instance that has the potential to tamper with the data a CPU or GPU is processing on behalf of a different cloud customer. Saileshwar said that Amazon Web Services and smaller providers such as Runpod and Lambda Cloud all provide A6000s instances. (He added that AWS enables a defense that prevents GPUhammer from working.)

Not your parents’ Rowhammer

Rowhammer attacks are difficult to perform for various reasons. For one thing, GPUs access data from GDDR (graphics double data rate) physically located on the GPU board, rather than the DDR (double data rate) modules that are separate from the CPUs accessing them. The proprietary physical mapping of the thousands of banks inside a typical GDDR board is entirely different from their DDR counterparts. That means that hammering patterns required for a successful attack are completely different. Further complicating attacks, the physical addresses for GPUs aren’t exposed, even to a privileged user, making reverse engineering harder.

GDDR modules also have up to four times higher memory latency and faster refresh rates. One of the physical characteristics Rowhammer exploits is that the increased frequency of accesses to a DRAM row disturbs the charge in neighboring rows, introducing bit flips in neighboring rows. Bit flips are much harder to induce with higher latencies. GDDR modules also contain proprietary mitigations that can further stymie Rowhammer attacks.

In response to GPUhammer, Nvidia published a security notice last week reminding customers of a protection formally known as system-level error-correcting code. ECC works by using what are known as memory words to store redundant control bits next to the data bits inside the memory chips. CPUs and GPUs use these words to quickly detect and correct flipped bits.

GPUs based on Nvidia’s Hopper and Blackwell architectures already have ECC turned on. On other architectures, ECC is not enabled by default. The means for enabling the defense vary by the architecture. Checking the settings in Nvidia GPUs designated for data centers can be done out-of-band using a system’s BMC (baseboard management controller) and software such as Redfish to check for the “ECCModeEnabled” status. ECC status can also be checked using an in-band method that uses the system CPU to probe the GPU.

The protection does come with its limitations, as Saileshwar explained in an email:

On NVIDIA GPUs like the A6000, ECC typically uses SECDED (Single Error Correction, Double Error Detection) codes. This means Single-bit errors are automatically corrected in hardware and Double-bit errors are detected and flagged, but not corrected. So far, all the Rowhammer bit flips we detected are single-bit errors, so ECC serves as a sufficient mitigation. But if Rowhammer induces 3 or more bit flips in a ECC code word, ECC may not be able to detect it or may even cause a miscorrection and a silent data corruption. So, using ECC as a mitigation is like a double-edged sword.

Saileshwar said that other Nvidia chips may also be vulnerable to the same attack. He singled out GDDR6-based GPUs in Nvidia’s Ampere generation, which are used for machine learning and gaming. Newer GPUs, such as the H100 (with HBM3) or RTX 5090 (with GDDR7), feature on-die ECC, meaning the error detection is built directly into the memory chips.

“This may offer better protection against bit flips,” Saileshwar said. “However, these protections haven’t been thoroughly tested against targeted Rowhammer attacks, so while they may be more resilient, vulnerability cannot yet be ruled out.”

In the decade since the discovery of Rowhammer, GPUhammer is the first variant to flip bits inside discrete GPUs and the first to attack GDDR6 GPU memory modules. All attacks prior to GPUhammer targeted CPU memory chips such as DDR3/4 or LPDDR3/4.

That includes this 2018 Rowhammer variant. While it used a GPU as the hammer, the memory being targeted remained LPDDR3/4 memory chips. GDDR forms of memory have a different form factor. It follows different standards and is soldered onto the GPU board, in contrast to LPDDR, which is in a chip located on hardware apart from the CPUs.

Besides Saileshwar, the researchers behind GPUhammer include Chris S. Lin and Joyce Qu from the University of Toronto. They will be presenting their research next month at the 2025 Usenix Security Conference.

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.

Nvidia chips become the first GPUs to fall to Rowhammer bit-flip attacks Read More »

US warns companies around the world to stay away from Huawei chips

Huawei, NVIDIA, syndication, Tech / DJ Henderson / May 14, 2025

President Donald Trump’s administration has taken a tougher stance on Chinese technology advances, warning companies around the world that using artificial intelligence chips made by Huawei could trigger criminal penalties for violating US export controls.

The commerce department issued guidance to clarify that Huawei’s Ascend processors were subject to export controls because they almost certainly contained, or were made with, US technology.

Its Bureau of Industry and Security, which oversees export controls, said on Tuesday it was taking a more stringent approach to foreign AI chips, including “issuing guidance that using Huawei Ascend chips anywhere in the world violates US export controls.”

But people familiar with the matter stressed that the bureau had not issued a new rule, but was making it clear to companies that Huawei chips are likely to have violated a measure that requires hard-to-get licenses to export US technology to the Chinese company.

“The guidance is not a new control, but rather a public confirmation of an interpretation that even the mere use anywhere by anyone of a Huawei-designed advanced computing [integrated circuit] would violate export control rules,” said Kevin Wolf, a veteran export control lawyer at Akin Gump.

The bureau said three Huawei Ascend chips—the 910B, 910C, and 910D—were subject to the regulations, noting that such chips are likely to have been “designed with certain US software or technology or produced with semiconductor manufacturing equipment that is the direct produce of certain US-origin software or technology, or both.”

The guidance comes as the US has become increasingly concerned at the speed at which Huawei has developed advanced chips and other AI hardware.

Huawei has begun delivering AI chip “clusters” to clients in China that it claims outperform leading US AI chipmaker Nvidia’s comparable product on key metrics such as total compute and memory. The system relies on a large number of 910C chips, which individually fall short of Nvidia’s most advanced offering but collectively deliver superior performance to a rival Nvidia cluster product.

US warns companies around the world to stay away from Huawei chips Read More »

Trump admin to roll back Biden’s AI chip restrictions

AI, AI regulation, Biden adminstration, Biz & IT, Donald Trump, GPUs, Joe Biden, machine learning, NVIDIA, Policy, trump administration / Kelly Newman / May 8, 2025

The changing face of chip export controls

The Biden-era chip restriction framework, which we covered in January, established a three-tiered system for regulating AI chip exports. The first tier included 17 countries, plus Taiwan, that could receive unlimited advanced chips. A second tier of roughly 120 countries faced caps on the number of chips they could import. The administration entirely blocked the third tier, which included China, Russia, Iran, and North Korea, from accessing the chips.

Commerce Department officials now say they “didn’t like the tiered system” and considered it “unenforceable,” according to Reuters. While no timeline exists for the new rule, the spokeswoman indicated that officials are still debating the best approach to replace it. The Biden rule was set to take effect on May 15.

Reports suggest the Trump administration might discard the tiered approach in favor of a global licensing system with government-to-government agreements. This could involve direct negotiations with nations like the United Arab Emirates or Saudi Arabia rather than applying broad regional restrictions. However, the Commerce Department spokeswoman indicated that debate about the new approach is still underway, and no timetable has been established for the final rule.

Trump admin to roll back Biden’s AI chip restrictions Read More »

Nvidia GeForce xx60 series is PC gaming’s default GPU, and a new one is out May 19

AMD, gaming, GeForce, geforce rtx, geforce rtx 5060, intel, NVIDIA, Tech / DJ Henderson / May 6, 2025

Nvidia will release the GeForce RTX 5060 on May 19 starting at $299, the company announced via press release today. The new card, a successor to popular past GPUs like the GTX 1060 and RTX 3060, will bring Nvidia’s DLSS 4 and Multi Frame-Generation technology to budget-to-mainstream gaming builds—at least, it would if every single GPU launched by any company at any price wasn’t instantly selling out these days.

Nvidia announced a May release for the 5060 last month when it released the RTX 5060 Ti for $379 (8GB) and $429 (16GB). Prices for that card so far haven’t been as inflated as they have been for the RTX 5070 on up, but the cheapest ones you can currently get are still between $50 and $100 over that MSRP. Unless Nvidia and its partners have made dramatically more RTX 5060 cards than they’ve made of any other model so far, expect this card to carry a similar pricing premium for a while.

	RTX 5060 Ti	RTX 4060 Ti	RTX 5060	RTX 4060	RTX 5050 (leaked)	RTX 3050
CUDA Cores	4,608	4,352	3,840	3,072	2,560	2,560
Boost Clock	2,572 MHz	2,535 MHz	2,497 MHz	2,460 MHz	Unknown	1,777 MHz
Memory Bus Width	128-bit	128-bit	128-bit	128-bit	128-bit	128-bit
Memory bandwidth	448GB/s	288GB/s	448GB/s	272GB/s	Unknown	224GB/s
Memory size	8GB or 16GB GDDR7	8GB or 16GB GDDR6	8GB GDDR7	8GB GDDR6	8GB GDDR6	8GB GDDR6
TGP	180 W	160 W	145 W	115 W	130 W	130 W

Compared to the RTX 4060, the RTX 5060 adds a few hundred extra CUDA cores and gets a big memory bandwidth increase thanks to the move from GDDR6 to GDDR7. But its utility at higher resolutions will continue to be limited by its 8GB of RAM, which is already becoming a problem for a handful of high-end games at 1440p and 4K.

Regardless of its performance, the RTX 5060 will likely become a popular mainstream graphics card, just like its predecessors. Of the Steam Hardware Survey’s top 10 GPUs, three are RTX xx60-series desktop GPUs (the 3060, 4060, and 2060); the laptop versions of the 4060 and 3060 are two of the others. If supply of the RTX 5060 is adequate and pricing isn’t out of control, we’d expect it to shoot up these charts pretty quickly over the next few months.

Nvidia GeForce xx60 series is PC gaming’s default GPU, and a new one is out May 19 Read More »

Trump can’t keep China from getting AI chips, TSMC suggests

AI chips, Artificial Intelligence, china, Chips Act, Donald Trump, Huawei, NVIDIA, Policy, semiconductors, Taiwan Semiconductor Manufacturing Company, tariffs, TSMC, us-china trade war / Kelly Newman / April 22, 2025

“Despite TSMC’s best efforts to comply with all relevant export control and sanctions laws and regulations, there is no assurance that its business activities will not be found incompliant with export control laws and regulations,” TSMC said.

Further, “if TSMC or TSMC’s business partners fail to obtain appropriate import, export or re-export licenses or permits or are found to have violated applicable export control or sanctions laws, TSMC may also be adversely affected, through reputational harm as well as other negative consequences, including government investigations and penalties resulting from relevant legal proceedings,” TSMC warned.

Trump’s tariffs may end TSMC’s “tariff-proof” era

TSMC is thriving despite years of tariffs and export controls, its report said, with at least one analyst suggesting that, so far, the company appears “somewhat tariff-proof.” However, all of that could be changing fast, as “US President Donald Trump announced in 2025 an intention to impose more expansive tariffs on imports into the United States,” TSMC said.

“Any tariffs imposed on imports of semiconductors and products incorporating chips into the United States may result in increased costs for purchasing such products, which may, in turn, lead to decreased demand for TSMC’s products and services and adversely affect its business and future growth,” TSMC said.

And if TSMC’s business is rattled by escalations in the US-China trade war, TSMC warned, that risks disrupting the entire global semiconductor supply chain.

Trump’s semiconductor tariff plans remain uncertain. About a week ago, Trump claimed the rates would be unveiled “over the next week,” Reuters reported, which means they could be announced any day now.

Trump can’t keep China from getting AI chips, TSMC suggests Read More »