NVIDIA Hopper architecture

Transformer Engine with FP8.

Introduced
2022
Process
TSMC 4N
Named after
Grace Hopper
Products in database
3

Data center

ModelLaunchShader coresMemoryBandwidthPowerUS launch
H200
NVIDIA Hopper
2024—141 GB
HBM3e
4,800 GB/sup to 700 W—
GH200 Grace Hopper
NVIDIA Hopper
2023——4,900 GB/s450–1000 W configurable (CPU + GPU + memory)—
H100 SXM
NVIDIA Hopper
202216,896
CUDA
80 GB
HBM3
3,350 GB/sup to 700 W—