It wouldn't throw away any power if the GPU makers decided to 1. Enable SR-IOV (or whatever their equivalent is named) for all GPUs, so people could split the

It wouldn't throw away any power if the GPU makers decided to 1. Enable SR-IOV (or whatever their equivalent is named) for all GPUs, so people could split the GPU power as they wanted and allocate these shares to VMs 2. Continued to evolve SR-IOV to reduce overhead and allow more dynamic allocation of resources