Nvidia is making three new AI chips for sales into China that circumvent the latest US trade restrictions, according to a report released Thursday by industry analysts at Semianalysis.
Nvidia now has product samples for the new H20, L20 and L2 GPUs that will go into mass production in December, according to the analysts. They found that one of the new GPUs is 20% faster than an existing H100 in LLM inference performance.
Sales of the Nvidia H100 and A100 into China were banned by the US a year ago, but Nvidia produced alternatives that could still legally still be sold into China called the H800 and A800. On Oct. 17, the US Department of Commerce and its Bureau of Industry and Security announced it is also restricting export of Nvidia H800 and A800, among others, that could allow China to fuel breakthroughs in AI and computers deemed critical to Chinese military operations.
The US government is blocking exports of the fastest chips into China over worries about national security, including that they could be used by China to improve AI and sophisticated computers to build weapons that endanger the US and its allies. GPUs are sold by Nvidia, AMD and Intel but Nvidia has by far the latest share of the $7 billion market. Such chips are designed to accelerate the complicated computation needed to train AI models and also to provide inference in making AI actions, such as producing generative AI responses in chatbots like those from ChatGPT and others.
Semianalysis has reported that Nvidia has shipped hundreds of thousands of H800 and A800 chips to China, but the most recent US Commerce restrictions would also limit those sales because Commerce eliminated a provision on bidirectional chip-to-chip transfer rate that had allowed Nvidia to sell H800 and A800 chips.
RELATED: Update: Biden rules would hurt advanced chip sales to China by Nvidia, others
“We thought the US locked down every single loophole conceivable” with the latest restrictions, Semianalysis said in its report. “To our surprise, Nvidia still found a way to ship higher performance GPUS into China with their upcoming H20, L20 and L2 GPUs.”
Semianalysis added: “In short, Nvidia is perfectly straddling the line on peak performance and performance density with these new chips to get them through the new US regulations.”
An Nvidia spokesman said the company would not comment on the report or its accuracy. In October, Nvidia had said the tightened US Commerce restrictions were not expected to have a near-term meaningful impact on financial results and said it would continue to comply with all applicable regulations "while working to provide products that support thousands of applications across many different industries."
The Semianalysis report’s authors include analyst Dylan Patel, who has a strong reputation for accuracy and is well known in the industry. Patel has described the growing trade tensions between China and the US over advanced chips as the cold war of the 21st century. He told Fierce Electronics that the information about Nvidia’s latest H20, L20 and L2 chips came from a “Taiwan source,” without elaborating. Star Market Daily, a news outlet, also described the new Nvidia chips, citing unnamed sources .
US Commerce officials could not be reached for comment.
Chipmakers such as Nvidia and Intel depend heavily on sales into China for a variety of chips, not only GPUs, and the trade group Semiconductor Industry Assocation has raised concerns about trade restrictions on sales and investments in China. "Overly broad, unilateral controls risk harming the Us semiconductor ecosystem without advancing national security as they encourage overseas customers to look elsewhere," SIA said recently.
Semianalysis posted specifications on the newest chips, including that the H20 uses CoWoS, which is short for Chip on Wafer oon Substrate, a high-density packaging technology for high performance chips developed by TSMC in 2012. TSMC is widely regarded as the manufacturer of the most advanced chips, and produces chips on designs by Nvidia, among others.
The H20’s memory bandwidth is 4.0 Tbps, compared to the H100’s memory bandwidth of 3.4 Tbps, but the H100 is capable of nearly 20 times the teraFLOPS as the H20.