US administration reportedly about to release new rules seeking to prevent China from accessing more high-end AI GPU chips
The US is reportedly planning to take new measures aimed at further restricting China’s access to AI chips, following landmark export controls introduced in October 2022.
Such a move has been expected after companies including Nvidia and Intel introduced graphics processing units (GPUs) for AI processing tasks that complied with last year’s sanctions while delivering capabilities comparable to top-of-the-line processors.
Reuters cited an unnamed US official as saying the administration is likely to further restrict the speed at which GPUs can communicate with one another when networked together to carry out high-performance tasks such as AI training.
The previous rules capped AI GPUs at an input-output bandwidth of 600 GB/s, amongst other restrictions.
The new rules are expected as early as this week.
Nvidia in May released a version of its flagship H100 GPU for the Chinese market, called the H800, that complied with the restrictions, including a reduced interconnect speed.
Intel did the same, releasing a version of its Gaudi2 GPU for the Chinese market in June.
Reuters cited sources in the semiconductor industry as claiming the US administration has been seeking to block the H800 from China.
In June Nvidia said if the H800 and the related A800 were restricted the firm did not expect it would have an immediate material impact on its financial results.
The administration also plans to introduce a “performance density” parameter to help prevent future loopholes, the official reportedly said, without offering details.
In addition, the US reportedly plans to require firms to notify the government when they sell chips to China whose performance is just below the guidelines.
Those chips will be subject to a case-by-case national security review, but can be shipped unless the chipmaker is told otherwise.
A further addition to the rules is to aim at closing a loophole that gives Chinese units located overseas access to US AI chips.
Artificial intelligence has become a key battleground between the US and China this year, following on the success of OpenAI’s ChatGPT and skyrocketing investment in the field.