BEIJING, November 10 (TMTPost)— Nvidia Corporation, the U.S. semiconductor giant riding the artificial intelligence (AI) wave, seems to find a way to circumvent new export controls imposed by the Biden administration.
Nvidia has developed a new set of chips specially for China, and all of the new offerings-- HGX H2,L20 PCle and L2 PCle are modified varieties of its flagship H100 graphics processing unit (GPU), Chinastarmarket, a Chinese news outlet under the state media company Shanghai United Media Group, learned from sources in the industry on Thursday. Nvidia could release these GPUs as soon as November 16, and Chinese customers will take the delivery in the coming days at earliest, the sources said.
A report from the South China Morning Post the same day echoed Chinastarmarket’s news. The Hong Kong-based newspaper cited a distributor from mainland China that Nvidia has developed a new server, the HGX H20, and two new GPUs, the L20 and L2, for Chinese customers. The distributor revealed the new chips comply with the latest U.S. export curbs and will be delivered as replacements for the A800 and H800, two tailored version Nvidia specially designed for Chinese market.
TMTPost App learned from several industrial chain companies that Nvidia will soon launch new products under the name of HGX H20, L20 PCle, L2 PClel, based on Hopper architecture and Ada Lovelace architecture. These products will be rolled out as early as November 16, and Nivida will mass produce them from December to January. Unlike the previous reports, TMTPost learned on Friday that three new AI chips are not modified version but the weakened one. Compared with H100, the upcoming HGX H20 has limitations on bandwidth and computing speed, and its overall computing power is about 80% less than that of H100, namely, the comprehensive computing performance of H20 is equal to 20% of H100. In addition, the computing power cost of H20 is higher due to additional modules of HBM memory and NVLink interconnect.
Nvidia is the semiconductor designer that dominates the market for AI chips, which empower AI systems including the large language model behind ChatGPT. The company has modified some of flagship products including A100 and H100 for exports to China, including an alternative A800 chip, as the U.S. regulators last year banned it from selling its most advanced chips to China. But even A800, the weakened version of Nivida’s cutting-edge A100 processor, is not allowed for export without first obtaining a license according to the new restrictions effective last month.
The rule entitles “Implementation of Additional Export Controls: Certain Advanced Computing Items; Supercomputer and Semiconductor End Use; Updates and Corrections”, was introduced by the U.S. Department of Commerce on October 18, with a goal to limit China’s access to advanced semiconductors that could fuel breakthroughs in artificial intelligence and sophisticated computers. The rule was supposed to come into effect following a 30-day public comment period. But Nvidia disclosed on October 25 that the U.S. government informed the licensing requirements of a rule introduced last week, applicable to products having a “total processing performance” of 4800 or more and designed or marketed for datacenters, is effective immediately.
The recent reports were accurate, that would the latest sign of how important the Chinese market is for US semiconductor firms. A spokesman for Nvidia said late October that the company is seeking additional supply of advanced AI computing system, which use chips affected by the latest rule. The person said that the new export control will not have a meaningful impact in the near future.
The Wall Street Journal reported right after the spokesperson’s remark that Nvidia could lose orders worth of $5 Billion in China due to the latest chip export control. The report said Nvidia could be forced to cancel next year orders from Chinese buyers. It suggested the company may face a hit of at least $5 billion as orders from major Chinese companies next year exceeded $5 billion. Chinese leading AI and cloud-computing companies including Alibaba Group, TikTok parent ByteDance, Baidu had made large orders for delivery next year, according to the report.