News on September 29, according to Bloomberg, as part of efforts to strengthen the local semiconductor industry and respond to United States sanctions, China is encouraging local manufacturers to buy domestic artificial intelligence (AI) chips to replace Nvidia's products.
Chinese regulators have been discouraging companies from buying Nvidia's H20 chips, according to people familiar with the matter, though the policy has taken a guiding step rather than an outright ban to avoid hindering the development of homegrown AI startups and raising tensions with the United States.
Analysts say the move is aimed at helping China's local AI chipmakers gain more market share while helping local tech companies prepare for possible further restrictions in the United States. China's leading AI processor manufacturers include Cambrian and Huawei. Earlier this year, there were also rumours in the industry that there was also guidance asking local EV makers to source more chips from local chipmakers as part of China's campaign to achieve self-sufficiency in key technologies.
Some companies will continue to use H20 chips and scramble to buy more H20 chips before the United States is expected to impose sanctions at the end of this year in order to quickly improve their AI capabilities, although they are also buying Huawei's own chips, a person familiar with the matter said.
Because China wants local companies to build the best AI systems. So, according to people familiar with Chinese's AI policy, this means that local companies will still tolerate the practice if they need to buy some foreign semiconductors instead of domestic alternatives.
Nvidia H20 is the strongest AI chip among the three China-specific AI acceleration chips that meet the requirements of United States export control after the previous A100/A800 and H100/H800 were banned from being exported to China. Although it is a "castrated" version based on the H100, it can provide up to 296 INT8 TOPS/FP8 TFLOPS of computing power, and also has 96 GB of HBM3 memory and 4.0 TB/s memory bandwidth, making it competitive with entry-level AI processors. However, compared with H100, its AI computing power is only less than 15% of that of H100, and some of its performance is even inferior to domestic AI chips, but its HBM capacity is higher than that of H100, which also makes it still have certain advantages over other domestic AI chips in actual AI training and inference.
After the H100 was banned, the Nvidia H20 proved to be a good replacement, attracting a lot of interest from Chinese tech giants. Most Chinese AI companies currently have developed their application ecosystems on NVIDIA's CUDA platform, so transitioning to AI chips from other companies will be expensive and time-consuming. Given that the H20 GPU is fully compatible with NVIDIA's CUDA platform, it is the preferred choice for many companies and applications, although it is much slower than the fully functional H100.
Previously, Dylan Patel, an analyst at SemiAnalysis, said that although the performance of the H20 is lower than that of some Chinese domestic AI chips "on paper", thanks to its integrated large-capacity HBM memory, Nvidia H20 is still "a little ahead" in actual AI applications. At the same time, it expects Nvidia to ship more than 1 million new H20 chips to the Chinese market this year, each of which will cost between $12,000 and $13,000, and will bring more than $12 billion in revenue to Nvidia.
However, Jefferies pointed out in a report in July that Nvidia H20 products could face a sales ban when the United States re-examines its semiconductor export policy in October. The ban can take many forms, including a product-specific ban, reducing the computing power of a chip, or limiting its storage capacity.
Nvidia CEO Jensen Huang said on Friday local time that he is doing everything he can to serve Chinese customers and comply with the requirements of the United States government restrictions. "The first thing we do is comply with whatever policies and regulations are being implemented," Huang said in an interview with Bloomberg TV. We have a lot of dependent customers there and we will do our best to support them. "
Despite the many restrictions in the United States, China's AI industry is indeed thriving. ByteDance and Alibaba have been aggressively investing, while a slew of startups are vying for leadership. When it comes to developing large language models (the key technology behind generative AI), there are six well-known players:01. AI, Baichuan, Moonshot, MiniMax, Stepfun, and Zhipu.
