It reaches parity because CUDA's hot path isn't forwarded at all. nvkvm forwards the ioctl control surface: context and channel setup, allocation, mapping and stops there. The guest maps the real device memory, so BAR writes from userspace go straight to hardware, and launching a kernel is a write to memory the guest already has in its VA space.