What Is Grace Blackwell?_
The Grace Blackwell platform pairs NVIDIA’s Grace CPU (based on ARM Neoverse cores) with a Blackwell GPU in a single module, sharing memory and bandwidth over a coherent NVLink-C2C link. This architecture is designed for AI and HPC workloads that benefit from tight CPU-GPU coupling, such as large-scale model training and inference. The platform appears in both single-module and multi-module configurations, often as the building block for larger systems like the GB200 NVL72.
Technical Details
The Grace CPU and Blackwell GPU are connected via NVLink-C2C, a chip-to-chip interconnect that provides coherent memory access and high bandwidth, typically in the hundreds of GB/s. The platform supports a large unified memory capacity per module, with the GPU accessing the CPU’s LPDDR5X memory directly. In data-center deployments, Grace Blackwell modules are liquid-cooled and interconnected via copper NVLink for GPU-to-GPU communication, while scale-out networking uses separate InfiniBand or Ethernet over fiber. The exact power and thermal specifications are defined by the OEM and NVIDIA’s reference design.
How Leviathan Systems Works with Grace Blackwell
In our field work, Grace Blackwell modules are installed as pre-assembled trays in the rack, with liquid cooling loops connected to the cold plates and NVLink cables routed between modules. We verify the NVLink-C2C link status during commissioning using NVIDIA’s diagnostic tools.
Related service
GPU Rack Assembly & Rack-and-Stack →