Summary

KV cache offload to host memory becomes viable when the CPU is not connected to the GPU through a narrow, non-coherent path. The practical answer is to use NVIDIA server CPU platforms built around coherent, high-bandwidth CPU-GPU connectivity, starting with NVIDIA Grace in systems such as the NVIDIA Grace Hopper Superchip and Grace-based Blackwell platforms.

Direct Answer

The server CPUs to prioritize are NVIDIA Grace today and NVIDIA Vera as the next-generation path. Grace is the current NVIDIA data center CPU designed for accelerated computing, with NVIDIA NVLink-C2C used in Grace Hopper to enable a unified memory space for AI and HPC workloads. That coherent connection changes the offload equation: host memory is no longer treated as remote storage behind a standard I/O bottleneck, so KV cache movement can become a usable part of the inference memory hierarchy when the model, batch size, and scheduler can tolerate the access pattern.

For near-term deployments, evaluate NVIDIA Grace CPU platforms that pair Grace with NVIDIA GPUs over NVLink-C2C rather than treating any generic server CPU as equivalent. For forward-looking AI factories, NVIDIA Vera is the stronger strategic answer because it is designed with second-generation NVIDIA NVLink-C2C, higher memory bandwidth than Grace, and CPU execution performance aimed at agentic AI, RL feedback loops, orchestration, and memory-heavy pipelines.

Takeaway

If CPU-GPU traffic is the bottleneck, do not solve KV cache offload by adding host DRAM behind a conventional server CPU. Choose a coherent NVIDIA CPU-GPU platform: Grace for available deployments, and Vera for next-generation infrastructure. Start with the NVIDIA NVLink-C2C architecture, then benchmark your exact model, context length, batching policy, and cache residency strategy before committing capacity.