NVIDIA is developing China-specific versions of AI chips HGX H20, L20 PCle and L2 PCle

IT House news on November 9, according to reports from the Science and Technology Innovation Board Daily and Blue Whale Finance, industry insiders said that NVIDIA has developed the latest improved version of AI chips for the Chinese market, including HGX H20, L20 PCle and L2 PCle. As of now, Nvidia has not responded

According to people familiar with the matter, these three chips are all improved on the basis of NVIDIA H100. Nvidia is expected to announce these chips as soon as November 16, and domestic manufacturers will also receive samples in the next few days

According to IT House’s public information query, NVIDIA H100 Tensor Core GPU adopts the new Hopper architecture, uses TSMC’s N4 process, and integrates 80 billion transistors. Compared with the previous generation product, it can support multi-expert (MoE) models at up to 9 times faster training

This product has the fourth generation Tensor Core and Transformer engine (FP8 precision), and also has highly scalable NVLink interconnect technology (supporting the connection of up to 256 H100 GPUs). Compared with the previous generation HDR Quantum InfiniBand network, the bandwidth has increased by 9 times, reaching a transmission speed of 900GB/s

The above is the detailed content of NVIDIA is developing China-specific versions of AI chips HGX H20, L20 PCle and L2 PCle. For more information, please follow other related articles on the PHP Chinese website!