Open-source prototype shares Nvidia GPU across four virtual machines

The developer of virtio-nvgpu has published an open-source prototype that lets Linux virtual machines use an Nvidia graphics card retained by the host. In tests reported by the developer, four guests rendered concurrently on one RTX 3060, with a combined frame rate of 103.7 frames per second against 102.9 for one guest under the same load. The software forwards Nvidia driver requests between a guest and host while the guest's own driver builds graphics commands. Its target is remote video streaming, where each guest renders and encodes frames on the GPU. The measurements cover a synthetic 720p workload; CUDA has not been tested beyond device enumeration, and the planned per-guest isolation component is unfinished.