NVIDIA's H100 accelerator is renowned for its powerful ability to support AI workloads, making it a highly coveted product. However, exporting US-made products to certain countries such as China can be difficult due to export regulations. 

To solve this issue, NVIDIA has developed a specific version of its H100 GPU called the H800 model for the Chinese market. The H800 features a reduced chip-to-chip interconnect speed of 300 GB/s, which is lower than the H100's speed. Similarly, last year, NVIDIA developed a China-specific version of its A100 model, labeled A800, with a reduced interconnect bandwidth of 400 GB/s.

According to NVIDIA, the H100 delivers up to 9 times faster AI training performance and up to 30 times speedier inference performance than the previous A100 model. This high level of performance makes the H100 a highly desirable product. To comply with export rules, Reuters reports that NVIDIA has modified the H100 to sell it as the H800 in China. Similarly, NVIDIA trimmed the interconnect speed of the A100 from 600 GB/s to 400 GB/s and rebranded it as the A800 to commercialize it in China.

It is unclear whether the CUDA or Tensor core count has been affected by the reduction in interconnect speed. However, the lower bandwidth is likely to impact latency and slow down the workload, especially when training large models. This is because transferring large amounts of data between chips takes time. NVIDIA has confirmed that the 800 series products comply with export control regulations but has not provided any further details about differences between the H800 and the H100. Despite Reuters' efforts to obtain more information from an NVIDIA spokesperson, the representative merely stated that "our 800-series products are fully compliant with export control regulations."

Nvidia modifies its H100 Hopper GPU and markets it as the H800 to Chinese market