Hopper vs Blackwell

Hopper and Blackwell are two generations of NVIDIA data center GPU architecture, with Hopper behind the H100 and H200 and Blackwell behind the B200 and B300. A Hopper GPU is one die with 80 billion transistors. A Blackwell GPU joins two reticle-limited dies over a 10 TB/s link into one GPU with 208 billion transistors. The B300 uses Blackwell Ultra, an upgraded two-die Blackwell.

Hopper tops out at 80 GB of HBM and 3.35 TB/s on the H100, and 141 GB of HBM3E and 4.8 TB/s on the H200. Blackwell holds up to 192 GB and Blackwell Ultra up to 288 GB, both at 8 TB/s. NVLink doubles from 900 GB/s per GPU to 1,800 GB/s, counting both directions. Dense FP8 rises from 2 PFLOPS to up to 5 PFLOPS. Blackwell introduced NVFP4, a 4-bit floating point format, at up to 10 PFLOPS dense. Maximum power climbs from 700 W to 1,200 W, or 1,400 W on Blackwell Ultra.

Choose Hopper when the model fits in an H200's 141 GB, runs at FP8 or FP16, and the racks are built for 700 W GPUs. Choose Blackwell when the model needs FP4, more memory per GPU, or faster NVLink, and the facility can feed up to 1,200 W per GPU.

Sources

Last verified October 7, 2026.