[throwaway84742]

And large number of cores comes in handy as well for the more efficient neural network architectures with complex data augmentation. I have several such networks where 16-core Threadripper is a bit of a bottleneck on a 4-GPU machine. We do the usual stuff: random scaling, cropping, rotation, noise, etc. Super easy, parallelizable stuff. That bottleneck will get worse once Volta consumer cards start showing up.