NVIDIA H800

What the H800 is

The H800 is a data-center GPU that NVIDIA made for the Chinese market as a version of its flagship H100, adapted so that it could be exported to China under the US export rules issued in October 2022. Reuters reported the chip on 21 March 2023 and said it was already in use by the cloud units of Alibaba, Baidu and Tencent^[1].

An earlier workaround

The H800 was not NVIDIA's first China-specific part. It followed the A800, an earlier China version of the A100 in which some capabilities had been reduced; Reuters reported that chip in November 2022^[1]. Both answered the same problem. The October 2022 rules did not ban data-center chips outright; they targeted chips that combine powerful computing with high chip-to-chip transfer rates^[2]. Weakening the links between chips was a way to stay inside the rules while keeping most of the computing power.

What changed

NVIDIA declined to say how the H800 differs from the H100, saying only that its 800 series products are fully compliant with export control regulations. A chip-industry source told Reuters that the main change was to cut the chip-to-chip data transfer rate to about half that of the H100^[1]. In a training cluster, that rate sets how quickly GPUs can exchange data, and the more a model is split across chips, the more often communication becomes the bottleneck. Trading interconnect for compliance therefore means less efficient work across chips even when the raw computing power is similar.

The rules tighten

On 17 October 2023 a tightened US rule was announced. It applied to data-center chips with a "total processing performance" of 4800 or more, and it covered the H800 and the A800 as well^[2]. On 23 October 2023 NVIDIA said the US government had made the new controls effective immediately for the A100, A800, H100, H800 and L40S^[3].

Inside DeepSeek-V3

DeepSeek-V3, released in December 2024, was trained on a cluster of 2,048 NVIDIA H800 GPUs^[4]. Each node holds eight GPUs connected through NVLink and NVSwitch, and the nodes are linked by InfiniBand. The report gives NVLink bandwidth inside the cluster as 160 GB/s, about 3.2 times InfiniBand's 50 GB/s, and to reduce cross-node traffic it limits each token to at most four nodes^[4].

The report states that full training took 2.788 million H800 GPU hours — 2.664 million of them for pre-training, which took under two months — or $5.576 million at an assumed $2 per GPU hour. Each trillion tokens took 180,000 GPU hours, or 3.7 days on the cluster. The report stresses that these figures cover only the final training run, not the earlier research and ablations^[4].

The DeepSeek team also described working within the chip's limits. The H800 has 132 streaming multiprocessors, of which they reserved 20 for communication, and they observed that FP8 matrix-multiply accumulation on the H800 keeps only about 14 bits of precision. From these constraints they drew up hardware design suggestions for AI chip vendors^[4].

Where it sits

The H800's public history is short: it was reported in March 2023 as already running in the cloud units of three Chinese companies^[1], and in October 2023 it was covered by the tightened rules, effective immediately^[2][3]. It marks how export controls changed which chips Chinese AI teams could buy, and the DeepSeek-V3 report preserves an unusually detailed account of how one such team organised training around the hardware it had — down to interconnect bandwidth, accumulation precision and parallelism — and turned those limits into design requests for chip makers^[4].