Primitives should probably be able to tell: * can they use GPUs to run * do they require GPUs to run * how many GPUs do they need, if multiple are available? Moreover, there should be some interface to tell which GPU a primitive should use. Or is this caller's responsibility to mask a particular GPU? Sidenote: we should also standardize which approach for accessing GPUs do we want to use in our Docker images. How to run that on Kubernetes?