What I miss most in QubesOS is ability to connect my GPU hardware to a specific VM (or app, or container) the way I do the mic, camera, and USB peripherals. I don't need it for rendering the browser GUI, where the CPU is wholly adequate, but it would be helpful for e.g. mpv, and as a back-end for OpenCL/OpenMP programs. This seems like it ought to be much easier than trying to share such a GPU with multiple VMs simultaneously.

The problem seems to be how to do it securely. You could imagine faking GPU hardware for the VM image to talk to. But maybe Vulkan would be a better API? Has any work been done on a wrapper forwarding calls that refer to VM-mapped addresses to a native Vulkan? (What happens in WSL2?)

I find Kompute ( https://github.com/KomputeProject/kompute , https://www.youtube.com/watch?v=Xz4fiQNmGSA ) that seems to be a serious attempt at building general computation infrastructure on top of Vulkan. Is exposing a Vulkan wrapper, and building compute kernels on top of that, a practical avenue?

This all seems like something someone must have throught through already, but I have not found anything published about it.