According to the news, Nvidia plans to launch a special version of the H20 chip to the Chinese market in Q2

IT Home reported on January 9 that, according to Reuters, Nvidia plans to start mass production of AI chips for China in the second quarter of 2024 to comply with the new U.S. export control regulations, and the strongest model is H20. Nvidia declined to comment.

In addition to the HGX H20, Nvidia also plans to launch two other PCIe chips that comply with US restrictions - L20 and L2, but the details are not known.

According to sources, H20 was originally scheduled to launch in November last year, but the plan was delayed due to "server manufacturers having problems integrating the chip." One of the people familiar with the matter said that the initial production capacity is very limited, and Nvidia needs to meet orders from mainstream customers first.

According to SemiAnalysis, H20, L20, and L2 contain most of the features that NVIDIA uses for AI work, but in order to comply with the new rules, the three new products have been cut in terms of computing power, but they are still more than 20% faster than H100 in LLM inference.

IT Home inquired about public information and learned that the NVIDIA H100 Tensor Core GPU uses the new Hopper architecture, based on TSMC's N4 process, and integrates 80 billion transistors. Up to 9x faster training for multi-expert (MoE) models compared to the previous generation.

It features fourth-generation Tensor Cores and a Transformer engine (FP8 precision), as well as highly scalable NVLink interconnect technology that can connect up to 256 H100 GPUs and deliver up to 9x more bandwidth than the previous generation with HDR Quantum InfiniBand networking at 900GB/s.