It allows you to run CUDA and graphics in VMs without VFIO/vGPU by forwarding the NVIDIA driver's own ioctl surface from the guest to the host.
That means stock libcuda and real Vulkan ICD run in the guest - no CUDA/Vulkan API remoting or translation. Display, LLM, CUDA, headless supported. Your desktop keeps using the same card, multiple VMs supported.
Happy to go into the details of the architecture.
FranciscoCarlos•21m ago