virtio-nvgpu forwards Nvidia driver ioctls to give a KVM guest near-native GPU access

Sources: primary, discussion
Summary: virtio-nvgpu translates at the Nvidia driver ioctl ABI rather than at the graphics API, so the guest builds its own command buffers. The project measures about 0.02 host crossings per frame against thousands for Venus, and states that guest-side NVENC works. The README states there is no IOMMU boundary, which leaves the host Nvidia driver inside the trusted computing base.
Why it matters: Driver-ABI forwarding gets a KVM guest near-native GPU throughput and working NVENC, at the cost of a trust boundary that VFIO passthrough keeps.
Follow-up: Track whether the project adds an IOMMU boundary or an upstream submission, and which driver versions the ABI forwarding tracks.