Nestrilabs virtio-nvgpu delivers near-bare-metal GPU performance in KVM virtual machines

Nestrilabs released virtio-nvgpu, an open-source virtualization technology that allows KVM virtual machines to access NVIDIA GPUs with performance nearly identical to bare-metal systems. This shift matters because it removes the traditional performance penalty associated with GPU sharing in virtualized environments. Users running multiple workloads can now utilize native driver stacks without the overhead of API translation layers.

Open-source driver forwarding cuts virtualization overhead to under 2 percent

The project introduces an innovative driver command forwarding mechanism that bypasses standard translation bottlenecks. This approach enables unmodified native driver stacks, including Vulkan and NVENC, to operate directly within the virtual machine. The solution currently supports Linux systems running NVIDIA driver version 535.129.03 or higher.

• Performance Overhead: Less than 2% frame time loss
• Test GPU: NVIDIA RTX 3060
• Virtualization Technology: virtio-nvgpu
• Minimum Driver Version: 535.129.03
• Test Resolution: 720p

Testing on an NVIDIA RTX 3060 system demonstrates that virtual rendering frame time loss remains under 2 percent. In multi-tenant scenarios, four virtual machines running identical loads achieved a total of 103.7 frames per second. This total slightly exceeds the 102.9 frames per second recorded in single-VM tests, a variance attributed to measurement error margins.

The technology currently operates only at 720p resolution, as validation has been limited to this specific setting. Benchmarks involving approximately 813,000 frames showed an average crossing of the virtual machine boundary every 59 frames. These metrics confirm that the command forwarding mechanism effectively minimizes context switching overhead.