Nestrilabs published virtio-nvgpu, an experimental virtio device enabling near-native NVIDIA GPU access inside KVM virtual machines. A guest renders within 2% of bare metal performance while costing the same CPU. The project forwards NVIDIA kernel driver ioctls between a Linux guest and host at the driver ABI level, bypassing API-level translation entirely. The guest runs NVIDIA's own unmodified user-mode drivers.

The target use case is headless streaming: a compositor inside the VM renders, composites, and encodes frames on the GPU, then sends compressed video out. Measured on an RTX 3060 with driver 595.99.02, a guest matches bare metal within noise for frames above 2 ms. CPU usage is identical — nothing is spent on forwarding in the render loop.

Four guests on one card share evenly at 25.84, 26.49, 25.57, and 25.79 fps, totaling 103.7 against 102.9 for a single guest. The split is even to four decimal places. Four is what was tested, not a limit. CUDA is forwarded but untested beyond enumeration. The repository splits across GPL-2.0 for the guest kernel module and Apache-2.0 for the host-side Rust crate.