It is rumored that Nvidia's castrated version of the AI chip H20 has started pre-sale: the performance is not as good as Huawei, but the price is higher!
On February 2, according to Reuters, there have been market rumors recently that Nvidia's "castrated version" AI chip H20 series designed for the Chinese market has begun to accept pre-orders from channel providers, and the pricing is consistent with the products of its rival Huawei.
According to the report, Nvidia has set the price of H20's Chinese channel in recent weeks at between 12,000~15,000 US dollars, but some channel providers have quoted as high as about 110,000 yuan (equivalent to 15,320 US dollars). In comparison, the Huawei Ascend 910B is priced at around RMB 120,000.
According to a source, channel providers have also started selling servers with eight H20 GPUs assembled in advance at a price of 1.4 million yuan. In comparison, when a server with 8 H800s was released a year ago, it was priced at 2 million RMB. The channel provider told customers that H20 is expected to start a small number of shipments in the first quarter of 2024, and a large number of shipments will wait until the second quarter.
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.
Nvidia then began developing new AI chips for China, including the HGX H20, L20 PCle, and L2 PCle, all of which are modified from the Nvidia H100 to comply with the latest U.S. export control policies.
According to the previously exposed specifications, the strongest H20 among the three products is also a castrated version based on the 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 the GH100 chip, the performance density (TFLOPs/Die size) is only 2.9. In comparison, the H100 has 80GB of HBM3 memory, 3.4Tb/s of internal memory bandwidth, 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.
The Reuters report also said that unnamed sources revealed that some key areas of the H20 are not even as strong as the Huawei Ascend 910B.
Obviously, Nvidia's castrated version of the AI chip does not have any advantages compared with some local AI chips in China, which makes domestic AI manufacturers not very interested in it. Previously, it was reported that Alibaba and Tencent are transferring some advanced semiconductor orders to local companies and relying more on chips developed in-house. The same is true for Baidu and ByteDance.
Editor: Xinzhixun-Rogue Sword