I'm aware of SR-IOV. Widespread support would go a long way, but doesn't it require pre-slicing the GPU into discrete chunks? I want microVMs that can share share a GPU dynamically the same way they share overprovisioned CPU resources. Much more like containers.