drivers/vfio/pci/nvgrace-gpu · keep

NVIDIA Grace Hopper and Grace Blackwell GPU passthrough

Glue code that lets a hypervisor hand a whole NVIDIA Grace Hopper or Grace Blackwell GPU (GH200, GB200, GB300) directly through to a guest virtual machine. These are NVIDIA's flagship 2024-2025 data-center superchips used for large-scale AI training and inference, where cloud and HPC operators slice systems up with KVM/QEMU and assign GPUs to individual tenants.

keep conf=0.87 deploy=low replacement=none subsystem=vfio category=virtualization

87%

recommendation

It should stay in the kernel because it supports NVIDIA's current-generation Grace Hopper (GH200) and Grace Blackwell (GB200/GB300) AI superchips when assigning them to virtual machines. The code was added in 2024, gained GB300 device IDs in 2025, and is still receiving feature work, so it tracks hardware NVIDIA actively sells for data-center AI workloads. Deployments are niche today because this is enterprise AI infrastructure, not consumer gear, but the platform is clearly on an upward trajectory.

repository signals

3 files

1,292 source lines

21 commits, 5y

+1,432 / −127 lines added / removed, 5y

6 authors, 5y

monthly commits · 2021-04-21 → 2026-04-21 · 21 total · active in 9/61 months

sources

git.kernel.org
Driver was added in 2024 specifically for Grace Hopper GPU assignment via VFIO, indicating a new-generation platform rather than legacy carryover.
git.kernel.org
Upstream added GB300 device IDs in 2025, showing ongoing enablement for newer Grace Blackwell SKUs rather than retirement.
git.kernel.org
The directory received substantive maintenance in 2026, including poison-handling support, with no sign of removal-only churn.
nvidia.com
NVIDIA’s Grace Hopper product page says GH200 is currently available.
nvidia.com
NVIDIA markets GB200 Grace Blackwell systems as current products, confirming the hardware family remains in new deployments.

codex reasoning notes (technical)

Local shell inspection of Kconfig/main.c identified this as a real VFIO PCI driver for NVIDIA Grace Hopper/Blackwell GPU passthrough to KVM/QEMU. `lei` was unavailable and MCP lore tools were not exposed here, so upstream activity was established from local `git log` and mapped to canonical kernel.org commit URLs by canonical recall. NVIDIA product URLs were obtained via web search results. Evidence points to active development, new SKU additions (GB200/GB300), and current product availability; deployment is likely niche/enterprise AI virtualization rather than mass-market, so keep rather than deprecate.