Shulou(Shulou.com)11/24 Report--

CTOnews.com November 9 news, foreign science and technology media Semianalysis recently released a paid article, saying that one of the three Chinese special version of AI chips, in LLM reasoning, is more than 20% faster than H100.

CTOnews.com previously reported that Nvidia is about to launch HGX H20, L20 PCle and L2 PCle interface chips, all based on improvements to the Nvidia H100. Officials have produced these GPU samples and are expected to go into mass production next month.

Sources said that Nvidia is expected to be announced as soon as November 16, and domestic manufacturers will get samples in the next few days at the earliest.

This paid article details Nvidia H20, L20 and L2, including FLOPS data, NVLink bandwidth, power consumption, memory bandwidth, memory capacity, chip size, etc.

The article reads:

More importantly, one of the three Chinese special versions of GPU (it is not clear which one, but should mean H20) is 20% faster than H100 in LLM reasoning, and its performance is close to the new GPU released by Nvidia early next year.

Another point mentioned in the report is that Nvidia HGX H20 is in HGX format and TSMC's CoWoS package. The other two chips are mainly based on Ada, oriented to the game field, and do not require complex packaging.

The article also forecasts Nvidia's results for fiscal years 2024 and 2025, after it was reported that Nvidia had lost more than $5 billion worth of orders due to the increasing size of the US, and that the use of these Chinese-only GPU would improve Nvidia's performance.

Related readings:

"it is reported that Nvidia is developing Chinese special AI chips HGX H20, L20 PCle and L2 PCle."

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.