Earlier quoted context omitted.
I could be wrong but I think you can only pass through an entire GPU as long as it's unused on the host. Sharing a used GPU, in the same way we can share a CPU with no performance loss via Intel Hyper v, is not possible.
Sharing a CPU with no performance loss isn't possible either. The amount of CPU loss varies dramatically by the exact virtualisation stack used and by the application running on top of it, but it is very clearly observable.
Running it in Windows (host) I have a lower score than I get in VMWare running Ubuntu (guest).