hardware//NVIDIA//H100

The H100 is a GPU of the Hopper architecture, with a Transformer Engine that supports FP8 at the hardware level. It delivers 2 to 3× the training throughput of an A100 and costs about $25K to $40K per card.

The H100 is a GPU of the Hopper architecture, with a Transformer Engine that supports FP8 at the hardware level. It delivers 2 to 3× the training throughput of an A100 and costs about $25K to $40K per card.

Its export to China is restricted, which made it the GPU that turned chip policy geopolitical; NVIDIA answered with cut-down variants for that market (the H800, with a slower interconnect, and later the H20), and DeepSeek trained V3 on H800s. It is on every frontier AI lab's shopping list, and perpetually backordered.