NVIDIA H800 ☆

The H800 kept the H100's FP8 tensor throughput but cut NVLink bandwidth to about 400 GB/s and sharply reduced FP64 to satisfy the 2022 export rules. DeepSeek's V3 technical report describes training on a 2,048-GPU H800 cluster, engineering around the interconnect limits — the chip is now shorthand for how constrained hardware shaped Chinese frontier labs. Banned by the October 2023 rules.