The GB200 NVL72 combines 36 Grace CPUs and 72 Blackwell GPUs in a rack-scale, liquid-cooled design. Its 72-GPU NVLink domain operates as a single, massive GPU—enabling 30X faster real-time inference for trillion-parameter large language models (LLMs).
Critical to the NVIDIA GB200 NVL72, the GB200 Grace Blackwell Superchip pairs two high-performance NVIDIA Blackwell Tensor Core GPUs with an NVIDIA Grace™ CPU. The NVIDIA NVLink™-C2C interconnect facilitates direct connectivity between the Grace CPU and the two Blackwell GPUs.
