Reuters exclusively reported that just this week, the United States will introduce new regulations to further restrict GPU exports to China.
It is speculated that Nvidia’s “castrated version” H800, which is specially designed to cope with the restrictions, will also be hindered.
The most noteworthy thing is that the previous "bandwidth parameter" restriction plan will be replaced, and a series of strict new regulations including "performance density", communication rate, etc. are already in the pipeline. .
In addition, the United States will impose additional inspections on Chinese companies that try to purchase chips through other countries and block the acquisition of American AI from overseas subsidiaries chip.
At the same time, we continue to expand the number of Chinese chip design companies included in the "blacklist" and require overseas manufacturers to obtain U.S. licenses for orders.
It can be seen that the upcoming new measures may further weaken the ability of Chinese companies to develop AI.
At the end of August 2022, both NVIDIA and AMD stated that their advanced chips such as A100 and H100 have already Listed on the export control list by the U.S. Department of Commerce.
These chips have become the industry standard for the development of chatbots and other AI systems.
In order to cope with this limitation, NVIDIA launched a "customized version" of the A800 chip in November. Its performance is below the threshold set by the U.S. Department of Commerce as a replacement for the banned A100. Subsequently, the "customized version" H800 corresponding to the H100 was launched.
However, even though the H800 has significantly limited key performance, the US government believes that its computing power is still no less than that of the H100 in some cases.
In order to further strengthen export restrictions on AI chips, the United States plans to use a number of The new standard will replace the previous restrictions on the "Bandwidth Parameter" (Bandwidth Parameter), thereby completing the ban on H800 and subsequent chips.
It is said that the new regulations are more fatal for the training of AI models - greatly reducing the communication rate between AI chips.
You must know that it is impossible to train the largest AI model on one GPU, but many chips need to be tied together. If their communication rate is reduced, it will increase the difficulty and cost of AI development.
In addition, the United States also plans to introduce a parameter called "performance density" to prevent companies from coming up with any workarounds.
Insiders revealed that the United States will require companies to notify the Commerce Department before shipping semiconductor products to China that perform slightly below guideline requirements. The United States will decide whether these products can be exported based on specific circumstances.
Among them, chips used in consumer products such as notebooks are not subject to the new regulations. But manufacturers still need to inform the U.S. Commerce Department when ordering the most powerful consumer GPUs.
A report released by CSET in June 2022 stated that China's AI capabilities depend on whether it can obtain American chips - most are designed by American companies such as Nvidia, Xilinx, Intel and Microsemi.
The US government has also been working to block the channels for AI chips to enter China.
Reuters said that this rule update is likely to plug a loophole that allows Chinese companies to obtain U.S. AI chips through subsidiaries located overseas.
While the rules are not expected to include restrictions on cloud computing services, the United States is still seeking comment on the risks of such access and how it should be addressed.
The above is the detailed content of Castrate H800 again? The U.S. Department of Commerce's new policy tightens restrictions on GPU exports and is expected to be announced this week. For more information, please follow other related articles on the PHP Chinese website!