NVLink-C2C

NVLink-C2C is NVIDIA's chip-to-chip interconnect, a version of NVLink that joins a CPU and a GPU inside one superchip instead of linking GPUs across a rack. It lets an Arm-based Grace CPU and a GPU share a coherent memory model. NVIDIA says it spans PCB-level connections, multi-chip modules and interposers, and supports Arm AMBA CHI or Compute Express Link protocols. It is separate from the NVLink and NVSwitch fabric between GPUs.

In the GH200 Grace Hopper Superchip, NVLink-C2C provides 900 GB/s of coherent bandwidth, which NVIDIA puts at 7X PCIe Gen5, with up to 624 GB of combined CPU and GPU memory. The GB200 Superchip joins two Blackwell GPUs and a Grace CPU over it, and the Vera Rubin Superchip joins two Rubin GPUs and a Vera CPU. NVIDIA rates it at up to 6x the energy efficiency and 3.5x the area efficiency of a PCIe Gen 6 PHY.

NVLink-C2C is a platform choice, not an add-in option. On an x86 server with PCIe GPUs, the CPU-to-GPU path is PCIe. Pick a C2C system, such as GH200, GB200 or DGX Spark, when a model spills from GPU memory into CPU memory and that traffic matters.

Sources

Last verified October 30, 2026.