Grace-Blackwell

The NVIDIA GB200 combines the 72-core Grace CPU with the Blackwell GPU architecture to deliver 30x faster LLM inference performance.

The Grace-Blackwell Superchip (GB200) integrates the ARM-based Grace CPU with the high-performance Blackwell GPU via a 900GB/s bidirectional NVLink-C2C interconnect. This unified memory architecture eliminates traditional PCIe bottlenecks, enabling the system to handle massive 27-trillion-parameter models. By pairing 72 Blackwell GPUs in a single NVL72 rack configuration, the platform achieves a 25x reduction in total cost of ownership and energy consumption compared to the previous H100 generation. It is the definitive hardware standard for generative AI training and real-time inference at scale.

https://www.nvidia.com/en-us/data-center/gb200-nvl72/