Grace Blackwell superchip

A Grace Blackwell superchip is an NVIDIA processor that pairs an Arm CPU with one or two Blackwell GPUs and lets both address one coherent memory pool. In the GB200, the building block of the GB200 NVL72 rack, one Grace CPU links to two Blackwell GPUs over NVLink-C2C, NVIDIA's chip-to-chip interconnect. The GB300 swaps in Blackwell Ultra GPUs, and the smaller GB10 powers the DGX Spark desktop.

The GB200's Grace CPU has 72 Arm Neoverse V2 cores and up to 480 GB of LPDDR5X at up to 512 GB/s. Its two GPUs add 372 GB of HBM3E at 16 TB/s. In the DGX Station, a GB300 desktop version links one Blackwell Ultra GPU to a 72-core Grace over 900 GB/s NVLink-C2C. Its 252 GB of HBM3E and 496 GB of LPDDR5X form 748 GB of coherent memory. The GB10 instead uses 20 Arm Cortex-X925 and Cortex-A725 cores and up to 128 GB of LPDDR5X at 273 GB/s, within a 140 W thermal design power.

Pick a Grace Blackwell system when a model needs more memory than the GPU's HBM alone holds and the CPU's LPDDR5X can extend it, as in large-model inference. Pick an x86 host with PCIe or HGX GPUs when the software stack ships only x86 binaries.

Sources

Last verified October 7, 2026.