Chinese tech giant Tencent doesn鈥檛 seem to be affected by US export bans of computer chips that are crucial to the development of artificial intelligence systems 鈥� but even if such bans were more stringent, they may not be able to slow the country鈥檚 AI advancement, US research shows.
and his colleagues at the University of California, Berkeley, have analysed publications released by researchers at Tencent about the firm鈥檚 latest models, including its Hunyuan AI models. The team鈥檚 findings suggest that, in recent months, Tencent has publicly described using H100 and A100 graphical processing units (GPUs). These chips, made by US firm Nvidia, are commonly used in AI development because of their ability to handle multiple complex calculations simultaneously, and are key to the large language models (LLMs) that power AI such as ChatGPT.
Chinese use of these GPUs raises questions, the research suggests, because the US Department of Commerce introduced export bans on the H100 and A100 in October 2022 and , citing fears over how the Chinese military might use the chips to develop AI models. Despite this, Gupta and his colleagues have identified documentation where in their view Tencent appears to reference the use of those chips after those dates. It should be noted, however, that the controls do not place limits on the use of any chips already in China before the export ban.
Advertisement
鈥淲e have tested V100 and A100 GPUs,鈥� 鈥� the older V100 chips aren鈥檛 subject to export controls, however, it isn鈥檛 clear whether the A100 chips were obtained before or after export restrictions were introduced by the US, although there is no suggestion of wrongdoing or that there has been any contravention of export controls by Tencent. Tencent has stressed that it operates in compliance with all applicable laws.
Gupta鈥檚 work also shows that researchers believed to be connected to Tencent also appear to have used H100 chips as part of a project to generate realistic videos of video games using AI. These researchers鈥� work was described by a website that went live in September 2024 and has . It detailed an AI project called GameGen-O and lists five of seven authors as working for LightSpeed Studios, a division of Tencent Games. The authorship of the project was later 鈥� before the site was deleted entirely.
The GameGen-O website didn鈥檛 describe the GPUs used in the project, but in November, a different set of five of those seven authors created a new website GameGen-X, though no affiliation with Tencent is mentioned on this new site. Both websites use the same video examples to illustrate the research, suggesting they use the same hardware, and the two projects appear to be closely related, if not functionally identical.
In a paper about GameGen-X also , these authors say 鈥渁ll experiments are conducted on 8 H100 GPUs鈥�. The source of these GPUs and the timing of their use relative to the export ban is unknown, but there is no suggestion of illegal activity, and there is no suggestion that there has been any contravention of export controls by Tencent. The authors of the GameGen-X project didn鈥檛 respond to a request for comment, but a new version of the paper . The new version says 鈥渙ur training infrastructure consisted of 24 NVIDIA H100 GPUs鈥�. It is not known why the change was made.
Gupta and his team also analysed software code relating to the Hunyuan AI that has been publicly released by Tencent. Their research suggests there are signs that this code was designed to work with cutting-edge chips like the H100 and A100, such as mentions of a number format, , that isn鈥檛 compatible with older chips. They say that the Hunyuan code also appears to make , ruling out chips from other manufacturers.
All this means Gupta鈥檚 team believes that Hunyuan systems are designed to work with Nvidia鈥檚 latest chips. It should be noted that writing code in this way wouldn鈥檛 itself break export controls, because it is possible to create code without owning the chips to run it, and again there is no suggestion of wrongdoing.
Gupta鈥檚 view is that 鈥渋f export controls are working against Tencent, then they shouldn鈥檛 be able to train state-of-the-art LLMs. That鈥檚 the whole promise of these export controls.鈥� But because Chinese companies may have stockpiled chips before export controls were introduced, and because they have been able to eke out better performance using better software coding from lower-powered chips that aren鈥檛 covered by export controls, the ban has seemingly not achieved its goal of slowing down Chinese AI development.
Gupta thinks the US may need more wide-ranging restrictions to achieve its aims. 鈥淭hese export controls aren鈥檛 achieving their end goal,鈥� says Gupta. 鈥淭he one way to achieve their end goal is to ban every GPU ever to be sent to China.鈥�
Tencent declined to respond to specific questions about the use of banned GPUs in these projects, however it commented that 鈥淭encent has been transparent about its work in developing and training AI models, and we operate in full compliance with all applicable laws鈥�. The US Department of Commerce didn鈥檛 respond to a request to comment.
An Nvidia spokesperson says: 鈥淲hile export controls are important to ensure that the most advanced hardware systems are used responsibly, it鈥檚 entirely expected 鈥� and indeed beneficial 鈥� that developers worldwide continue to drive progress through new algorithms and open-source contributions, fostering overall industry advancement and delivering value to society as a whole.鈥�
at Princeton University says US export controls on semiconductor manufacturing tools 鈥� essentially, the tools that make the machines that make GPUs 鈥� are a more effective barrier to China鈥檚 AI development than outright blocks on the latest GPUs. 鈥淐hinese tech firms are throwing the kitchen sink at the AI compute problem,鈥� he says, meaning the need to have large computational resources to train AI. 鈥淭he controls on GPUs themselves may have diminishing returns given the difficulties of blocking new loopholes and workarounds.鈥�
Whether the US approach to AI will change when the country鈥檚 recently elected president, Donald Trump, takes power in 2025 is uncertain, but a general anti-China sentiment seems likely in the new administration. 鈥淚 would expect an even bigger yard and higher fence when it comes to China export controls,鈥� says Chan.
However, Gupta says simply banning chip exports won鈥檛 achieve the US goal of limiting Chinese AI development. 鈥淓xpert control should be a combination of software use cases and specific types of software combined with hardware,鈥� says Gupta. 鈥淪top looking at hardware spec sheets. Let鈥檚 look at actual performance. Let鈥檚 start the conversation as to what a more meaningful use-case based expert control strategy might look like.鈥�
arXiv
Article amended on 6 December 2024
We have clarified the research referred to and the scope of the export bans
