A GPU from NVIDIA designed for AI. In 2024, NVIDIA introduced the Blackwell B100 and B200, both of which are two dies linked together with 192GB of HBM3e memory and 208 billion transistors on a single chip. Named after the distinguished American mathematician David Blackwell, the B200 is roughly 25% faster than the B100 but is about 30 times faster than NVIDIA's H100 chip. See
H100,
A100,
GPU and
high bandwidth memory.
GB200 - Grace Blackwell Superchip
Superseding the Grace Hopper Superchip, the Grace Blackwell Superchip links two Blackwell B200 GPUs and one Grace CPU via an NVLink at 900 Gbps. The B200 GPU is 30 times faster than NVIDIA's H100 Tensor Core GPU and reduces energy up to 25 times. See
Grace Hopper Superchip.
NVL72 Datacenter Platform
Acting as a single GPU for datacenters, the NVL72 is a liquid-cooled rack-mounted system with 36 GB200s for a total chip count of 36 CPUs and 72 GPUs.
Not Your Basic GPU Chip
The cost of a Blackwell chip starts around $30,000, and hundreds are used in AI datacenters to process large language models. Each server rack can cost millions of dollars. See
AI datacenter and
rack mounted.
Blackwell Ultra and Rubin
Expected in 2025, Blackwell Ultra is an enhanced version of Blackwell with more memory. The successor to Blackwell is Rubin, expected in 2026. See
Rubin.