What Is Grace CPU?_
NVIDIA Grace is an ARM-based data center CPU designed to pair with NVIDIA GPUs via NVLink-C2C. Grace features 72 Neoverse V2 cores and uses LPDDR5X memory for high bandwidth and energy efficiency. The Grace CPU appears in GH200 (Grace Hopper), GB200, and GB300 systems. Its ARM architecture offers better performance-per-watt than x86 alternatives for GPU-paired workloads.
Technical Details
The Grace CPU is built on ARM's Neoverse V2 architecture with 72 high-performance cores. It uses LPDDR5X memory (up to 480 GB) instead of traditional DDR5, providing higher memory bandwidth per watt. The key innovation is NVLink-C2C, which connects Grace to its paired GPU at 900 GB/s bidirectional bandwidth — approximately 7x the bandwidth of PCIe Gen5. This eliminates the CPU-GPU memory bottleneck that limits traditional x86+GPU configurations. In GH200 systems, one Grace CPU pairs with one Hopper GPU. In GB200 and GB300 NVL72 systems, 36 Grace CPUs pair with 72 GPUs (2:1 GPU:CPU ratio). The ARM architecture means the software stack must support ARM builds, which is now standard for NVIDIA's AI software ecosystem.
How Leviathan Systems Works with Grace CPU
Leviathan Systems deploys Grace CPU-based systems (GH200, GB200, GB300) with the same infrastructure rigor applied to all GPU platforms, accounting for the different power efficiency profile of ARM-based systems in our deployment planning.