From GitHub · · 1m 10s
GitHub - nestrilabs/virtio-nvgpu
Share one NVIDIA GPU across multiple VMs at bare-metal speed with NVIDIA drivers.
In brief
Share one NVIDIA GPU across multiple VMs at bare-metal speed with NVIDIA drivers. Nestri's virtio-nvgpu lets users run NVIDIA GPU workloads in multiple microVMs without vGPU licenses or VFIO passthrough. Originally reported by GitHub.
What is virtio-nvgpu
Share one NVIDIA GPU across several VMs at bare-metal speed.
“Share one NVIDIA GPU across several VMs, at bare-metal speed.”
Why traditional methods fail
Passthrough locks the whole GPU to one VM. vGPU costs real money and only works on datacenter cards.
“Passthrough locks the whole GPU to one VM. vGPU costs real money and only works on datacenter cards.”
How it works
virtio-nvgpu forwards the NVIDIA driver instead of the graphics API. The guest builds its own GPU commands and writes them straight to GPU memory.
“virtio-nvgpu forwards the NVIDIA driver instead of the graphics API. The guest builds its own GPU commands and writes them straight to GPU memory”
Performance
On an RTX 3060: Within 2% of bare metal for any real game frame (>2 ms). Same CPU cost as running on the host.
“On an RTX 3060: Within 2% of bare metal for any real game frame (>2 ms)”
Who it's for
People who stream games or GPU apps from headless VMs (cloud gaming, remote desktops, CI with a real GPU, selfhosted "serverless" AI) on consumer NVIDIA cards.
“People who stream games or GPU apps from headless VMs (cloud gaming, remote desktops, CI with a real GPU, selfhosted "serverless" AI) on consumer NVIDIA cards.”
Before you use it
It isn't hardware isolation. The host NVIDIA driver is the trust boundary. If your tenants don't trust each other, use passthrough or vGPU.
“It isn't hardware isolation. The host NVIDIA driver is the trust boundary. If your tenants don't trust each other, use passthrough or vGPU.”
Here's the gist
Nestri's virtio-nvgpu lets users run NVIDIA GPU workloads in multiple microVMs without vGPU licenses or VFIO passthrough.
