Ultimately that's a driver issue. GPUs have had the equivalent of an IOMMU for almost two decades; longer than virtualization has existed in commodity hardware. But the heavily closed nature of the hardware, and it not really being a priority for the GPU vendors has led to it not getting the love it needs in that regard. You don't need to add vGPUs to get there.

Yeah, it's industry wide. We work directly with nvidia. They use our product. I thought the paper was interesting, definitely not a magic bullet for the problem. You used to be able to over provision with kubernetes 1.8 and before, but they fixed that useful bug for good reasons. I expect we'll have an official solution in Q1 next year, but I'm always on the look out to explore related implementations.