Grace-Blackwell GB10

The GB100 architecture integrates two Blackwell GPUs with a Grace CPU via a 900GB/s NVLink-C2C interconnect to deliver 20 petaflops of AI performance.

NVIDIA's Grace-Blackwell GB200 superchip marks a massive jump in compute density (30x faster LLM inference compared to H100). It pairs the ARM-based Grace CPU with two high-performance Blackwell GPUs, utilizing a second-generation transformer engine to handle 4-bit floating point (FP4) precision. This hardware is the backbone of the NVL72 rack system: a liquid-cooled powerhouse that acts as a single 72-GPU node with 130TB/s of aggregate bandwidth. By slashing energy consumption by 25x and reducing costs, it provides the massive scale required for trillion-parameter model training and real-time deployment.

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