US tackles loopholes in curbs on AI chip exports to China
A US official has confirmed that the United States is planning to implement measures aimed at preventing American chip manufacturers from supplying semiconductors to China that evade government-imposed restrictions. These actions are part of the Biden administration’s forthcoming efforts to further restrict the export of AI chips.
The forthcoming rules, which Reuters is disclosing for the first time, will be an extension of the comprehensive restrictions placed on the shipment of advanced chips and chipmaking equipment to China, initially unveiled in October of the previous year. While it’s anticipated that these updates will be introduced this week, it’s worth noting that such timetables can often experience delays.
The new regulations will specifically target some AI chips that currently fall just below existing technical specifications. Additionally, companies will be required to report their shipments of certain AI chips. This information was provided by an anonymous official. The US Department of Commerce, responsible for export controls, declined to comment.
This latest crackdown on tech exports to China occurs simultaneously with US efforts to improve relations with the world’s second-largest economy. Several high-ranking members of the Biden administration have engaged with their Chinese counterparts in recent months. However, the introduction of these new rules poses a potential complication to these diplomatic efforts.
The Biden administration’s rationale for these export restrictions is to prevent US chips and equipment from strengthening China’s military capabilities. In response, Beijing has accused the United States of abusing export controls to stifle Chinese companies. These restrictions represent a significant shift in US-China tech policy.
Also Read | Huawei’s (Kirin 9000S) New Mystery 7nm Chip from Chinese Fab Defies US Sanctions
The Chinese embassy in Washington has not yet responded to requests for comment.
In the past year, government restrictions prevented Nvidia, one of the most valuable chip manufacturers globally, from supplying two of its most advanced AI chips to Chinese customers. These chips have become the industry standard for developing chatbots and other AI systems. However, Nvidia quickly introduced new variants tailored for the Chinese market that were less advanced, effectively bypassing US export controls. One such chip, known as the H800, offered comparable computing power at certain settings used in AI work as their more powerful but restricted H100 chip. However, some key performance aspects were limited, according to a specification sheet seen by Reuters.
The United States now intends to establish new guidelines that will limit certain advanced data center AI chips that are currently not covered by existing restrictions. While the official did not specify which additional chips would be affected, sources have suggested that Nvidia’s H800 chip is one of the targets.
Nvidia, based in Santa Clara, California, has not provided an immediate response to requests for comment. In June, the company’s chief financial officer stated that if the H800 and a related chip called the A800 were restricted, it would not have an immediate material impact on the company’s financial results.
The new restrictions will not apply to chips intended for consumer products like laptops. However, companies will be required to inform the Commerce Department when fulfilling orders for the most powerful consumer chips to ensure they are not being used in ways that jeopardize national security.
To prevent excessively powerful AI chips from reaching China, the US plans to eliminate one of the parameters, known as the “bandwidth parameter,” that has been used to restrict the export of specific AI data center chips. By removing this parameter, another guideline will come into play, broadening the scope of covered chips. This is expected to result in a reduction in the speed at which AI chips can communicate with each other.
Also Read | A UK microchip company, Arm, plans to sell its shares in the United States
This is a significant development since the training of the largest AI models requires multiple chips working in unison. Slowing down their communication speed will make AI development more challenging and costly.
To deter potential workarounds, the US plans to introduce a “performance density” parameter. Further details on this parameter were not provided by the official.
These updated rules are designed to remain relevant as technology evolves. Companies will be required to notify the government about semiconductors with performance that falls just below the specified guidelines before shipping them to China. The government will assess on a case-by-case basis whether these chips pose a national security risk, but they can be shipped unless the chipmaker is informed otherwise.
These updates to the October 2022 rules may also close a loophole that allows Chinese companies to access American artificial intelligence chips through their overseas units, as previously reported by Reuters.
It is not expected that these rules will include restrictions on access to US cloud computing services or those of US allies. However, the US intends to seek feedback on the risks associated with such access and how they might be addressed.
Earlier this month, the Biden administration informed Beijing about its intentions to revise the contentious rules, as reported by Reuters. This move is part of the administration’s broader policy efforts aimed at promoting stability in the relationship between the United States and China.
