news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

On July 22, a report provided to clients by Jefferies Securities recently showed that the US government is considering new trade restrictions that could prevent Nvidia from selling its HGX-H20 AI GPU specifically for the Chinese market. If the restriction is officially implemented, Nvidia could lose about $12 billion in revenue.

Affected by the US export control policy, the export of NVIDIA's advanced AI chips to China has been restricted. The H20 is the most powerful of the three "castrated" GPUs launched by NVIDIA based on the H100 specifically for the Chinese market. It can provide up to 296 INT8 TOPS/FP8 TFLOPS computing power, and also has 96 GB 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 less than 15% of H100, and some performance is even inferior to domestic AI chips, but its HBM capacity is higher than H100, which also makes it still have certain advantages over other domestic AI chips in actual AI training and reasoning.

According to the latest forecast data from market research firm SemiAnalysis, AI chip giant NVIDIA will ship more than 1 million new NVIDIA H20 acceleration chips to the Chinese market this year. Each chip is expected to cost between US$12,000 and US$13,000, which is expected to bring NVIDIA more than US$12 billion in revenue.

But according to a recent report from Jefferies, Nvidia's H20 products may face a sales ban when the United States reviews its semiconductor export policy again in October this year. The ban may take many forms, including bans on specific products, reducing the computing power of chips or limiting their storage capacity.

Currently most Chinese AI companies have developed their application ecosystems on Nvidia's CUDA platform, so transitioning to others will be expensive and time-consuming. Given that the H20 GPU is fully compatible with Nvidia's CUDA platform, it is the first choice for many companies and applications, even though it is much slower than the full-featured H100.

Notably, despite export controls, Chinese companies have obtained advanced Nvidia GPUs for artificial intelligence and high-performance computing through intermediaries and by renting servers from companies such as Google and Microsoft.

The US could also expand export restrictions to other Asian countries such as Malaysia, Indonesia, Thailand, and potentially Chinese companies overseas. However, the report said that enforcing these expanded controls would be complex and challenging to implement effectively. Forcing (or denying) access to Chinese entities from the cloud could be problematic. Given the enormous processing performance requirements of large-scale language training for AI, it is almost inevitable that Chinese entities will have access to offshore GPU performance in some way, which means they will simply rent instances in the cloud.

Editor: Core Intelligence - Rurouni Ken