Hopper
NVIDIA data center GPU architecture unveiled in 2022, used in the H100 and H200. It introduced FP8 and the Transformer Engine.
Hopper is the GPU architecture NVIDIA unveiled in 2022 for data centers, named after the computer scientist Grace Hopper. It powers the H100 and H200, as well as the GH200 chip, which pairs a Hopper GPU with a Grace processor. It succeeded Ampere, the architecture of the A100, before being replaced in turn by Blackwell.
Hopper marked a turning point for generative AI. Its fourth-generation Tensor Cores handle FP8 precision, and its Transformer Engine automatically chooses between FP8 and FP16 depending on the model's layers, which significantly speeds up training and inference of Transformer-type models. It also brings fourth-generation NVLink, at about 900 GB/s per GPU, and HBM3 memory.
Between the H100 and H200, computing power is identical: the difference lies in the memory. The H100 SXM has 80 GB of HBM3, the H200 141 GB of HBM3e with a bandwidth of about 4.8 TB/s. When running large language models, memory is often the limiting factor, and the H200 then lets you load bigger models or handle more requests in parallel. On a product page, also check the form factor, SXM or PCIe, and the NVLink generation.
We offer H200-based configurations on quotation, with a reservation deposit refunded if you change your mind.