laitimes

The performance has shrunk by 85%, and the NVIDIA China special version of the AI chip can't be sold!

The performance has shrunk by 85%, and the NVIDIA China special version of the AI chip can't be sold!

The performance has shrunk by 85%, and the NVIDIA China special version of the AI chip can't be sold!

On January 8, according to foreign media reports, although Nvidia is preparing to launch a new customized "castrated version" of AI chips to the Chinese market to comply with the U.S. export control policy, Chinese manufacturers are not willing to actively buy.

Since November last year, Alibaba Group, Tencent and other large Chinese cloud computing companies have been testing samples of Nvidia's special chips, according to people familiar with the matter. They have indicated to Nvidia that the number of chips ordered from Nvidia this year will be far less than the banned Nvidia high-performance chips that they had planned to buy.

The new ban issued by the U.S. government on October 17, 2023, further restricts the export of Nvidia's high-performance AI chips, but before the ban officially takes effect on November 17, the U.S. government ordered Nvidia's export restrictions for "products with comprehensive performance of 4800 or more and designed or sold for data centers" to "take effect immediately." Nvidia also announced at the time that it would immediately stop shipping A100, A800, H100, H800 and L40S products.

Subsequently, according to industry chain sources, Nvidia is developing the latest improved version of AI chips for China, including HGX H20, L20 PCle and L2 PCle, which are all based on Nvidia H100 to comply with the latest U.S. export control policy.

The performance has shrunk by 85%, and the NVIDIA China special version of the AI chip can't be sold!

According to the previously exposed specifications, H20 is a "castrated version" of H100 GPU, with an internal memory capacity of 96 GB, a running speed of up to 4.0 Tb/s, and a computing power of 296 TFLOPs, using a GH100 chip, and a performance density (TFLOPs/Die size) of only 2.9. In comparison, the H100 has 80GB of HBM3 memory, 3.4Tb/s of internal memory, 1,979 TFLOPs of computing power, and a performance density of up to 19.4. In other words, the AI computing power of the "castrated version" of the H20 chip is less than 15% of that of the H100. Obviously, this performance does not meet the needs of Chinese AI manufacturers.

In the short term, NVIDIA's castrated AI chips do not have any advantages compared with some local AI chips in China, which makes domestic chips more and more attractive to buyers. Alibaba and Tencent are shifting some advanced semiconductor orders to local companies and relying more on chips developed in-house, people familiar with the matter said. The same is true for Baidu and ByteDance.

Editor: Xinzhixun-Lin Zi

Read on