GPUvm: Why Not Virtualizing GPUs at the Hypervisor?