NVIDIA's H20 AI Chip: Q2 Mass Production in China

On January 3, 2024, reports from WCCFtech indicate that NVIDIA is poised to recommence shipments of AI GPUs to China. In the second quarter of this year, production will commence for H20 and other AI GPUs, with the main boards being handled by the supplier, Wistron. This development comes after the U.S. government imposed new restrictions on the export of high-performance AI chips by NVIDIA on October 17, 2023. However, the export restrictions, applicable to products with "composite performance reaching 4800 or above and designed or sold for data center use," took immediate effect before the official enforcement of the ban on November 17.

In response to the restrictions, NVIDIA promptly ceased shipments of A100, A800, H100, H800, and L40S products. Following this, industry insiders revealed that NVIDIA is working on the latest improved version of AI chips specifically tailored for the Chinese market. These include HGX H20, L20 PCIe, and L2 PCIe, all derived from the enhanced H100 chip to comply with the latest U.S. export control policies.

According to previously disclosed specifications, H20 is a scaled-down version of the H100 GPU, featuring a 96 GB memory capacity, a high-speed of 4.0 Tb/s, and computational capabilities of 296 TFLOPs. It utilizes the GH100 chip, with a performance density (TFLOPs/Die size) of only 2.9. In comparison, H100 boasts 80GB HBM3 memory, a memory bandwidth of 3.4 Tb/s, and computational capabilities of 1,979 TFLOP, with a high performance density of 19.4.

Based on these calculations, H100 SXM is expected to be 6.68 times faster than H20 SXM (based on FP16 Tensor Core FLOPs). The GPU has a 400W TDP, an 8-way configuration in the HGX solution, retaining a 900 GB/s NVLINK connection, and offering 7-way MIG (Multi-Instance GPU) functionality.

Originally planned for release by the end of 2023, the launch of these chips has been delayed to early 2024 amid the escalating tensions in the U.S.-China semiconductor war. Presently, the U.S. appears to have softened its stance on sales to China, and NVIDIA assures that the chips will fully comply with the requirements and guidance of the U.S. Department of Commerce, facilitating the launch of the GeForce RTX 4090D in China.

On December 11, 2023, U.S. Secretary of Commerce Gina Raimondo mentioned in an interview that the Biden administration is in discussions with NVIDIA, considering limited sales of artificial intelligence chips to China. Raimondo emphasized that, given the commercial applications of most AI chips, NVIDIA should be allowed to sell them to China within certain limits. NVIDIA CEO Jensen Huang has consistently emphasized the company's active collaboration with the U.S. government to develop products that adhere to regulations, stating, "We intend to continue cooperating with the U.S. government and developing a new series of products in accordance with the new regulations."