The United States is expected to announce a new set of rules that aim to prevent Chinese firms from sourcing American-made high-end artificial intelligence (AI) chips through different channels.
An unnamed US official told Reuters that Washington plans to introduce new guidelines that will restrict certain advanced AI chips to China. Specifically, the US is mulling whether to ban exports of Nvidia’s H800 chip to China, according to the report.
The US will increase its efforts to check whether any Chinese firms are evading export restrictions by routing shipments through other nations, Bloomberg reported. It will also require overseas manufacturers to receive a US license before taking orders from Chinese chip design firms. The new rules will be published early this week, the reports said.
“China opposes US politicizing, instrumentalizing and weaponizing trade and technology issues,” said Wang Wenbin, a Chinese Foreign Ministry spokesperson.
“Arbitrarily placing curbs or forcibly seeking decoupling to serve political agenda violates the principles of market economy and fair competition, undermines the international economic and trading order, disrupts and destabilizes global industrial and supply chains and will eventually hurt the interests of the whole world,” he said.
Wang said China will closely follow developments and firmly safeguard its legitimate rights and interests.
Interchip bandwidth
In August last year, the Biden administration ordered US chipmakers to stop exporting graphic processors that operate at an interchip bandwidth of 600 gigabytes per second or above to China and Russia. Nvidia’s A100 and H100 chips and AMD’s MI250 chip are in the category of this new rule.
Nvidia later unveiled A800 and H800 processors, which work at 400 and 300 gigabytes per second respectively, specifically for Chinese markets. Both processors satisfy the US government’s export requirements.
In June this year, media reports said the Biden administration would soon ban the export of the A800 and H800 to China, though it has not yet announced the restrictions.
A US official told Reuters in a recent report that the bandwidth parameter will be replaced by a “performance density” parameter to help prevent future workarounds.
Companies can only ship to China AI chips with performance parameters that are below the export ban guidelines.
The Financial Times reported on August 9 that Chinese technology giants, including BAT (Baidu, Alibaba and Tencent) and ByteDance, have recently ordered US$5 billion of A800 chips from Nvidia.
The report said about 100,000 A800 chips worth $1 billion are scheduled for delivery this year while the remaining will be shipped in 2024.
Technology experts said the A800’s performance has reached 70% of the A100’s. It’s unclear whether the shipment of the A800 and H800 to China will be banned under the new guidelines.
“H800 is already a slower version of H100 but it may still be banned from being shipped to China,” a Beijing-based columnist said in an article published by AI era, a technology news website, on Monday. “The soon-to-come US curbs will hurt Chinese firms’ ability to develop their AI technologies.”
Import workarounds
Reuters reported in June that some retailers in a shopping mall in Huaqiangbei district in Shenzhen were found to have sourced small numbers of A100 chips and priced them at $20,000 each, double their usual price.
The Center for Security and Emerging Technology (CSET), a Washington-based think tank, said in a research report in June last year that the Chinese People’s Liberation Army (PLA) often buys commercial off-the shelf AI systems from Chinese academic institutions and private companies, which are not easily blocked by US restrictions on military end-users.
“Difficulties associated with tracking AI chips and the variety of potential vendors would make it challenging for US regulators to wage a targeted crackdown on the PLA’s intermediary chip suppliers,” said the CSET.
Of the 97 individual AI chips the CSET could identify in public PLA purchase records, nearly all of them were designed by Nvidia, Xilinx (now AMD), Intel or Microsemi. The CSET also said the PLA is placing orders for AI chips designed by US firms that are manufactured in Taiwan and South Korea.
In fact, sanctioned Chinese tech giant Huawei Technologies can still outsource the manufacturing of its Ascend 910 chip to Taiwan’s TSMC as the 7nm AI processor uses the company’s self-developed Da Vinci architecture.
According to a CITIC Securties research report published in July, more than 20 Chinese cities have used Ascend chips in their AI facilities while Huawei now has a 79% share in China’s AI computing center market.
At the World Artificial Intelligence Conference held in Shanghai on July 6, Zhang Dixuan, head of Huawei’s Atlas Data Center Business, said the company will boost the number of Ascend 910 cards in its Altas 900 Supercluster from 4,000 to 16,000 later this year.
Yao Yue, an IT writer, says such a cluster is enough to train a chatbot equivalent to GPT3.0.
Read: Chip war may thwart Shanghai plans to build AI hub
Read: What’s behind Samsung, SK Hynix chip war waivers?
Follow Jeff Pao on Twitter at @jeffpao3