A lot of deep learning workloads aren't necessarily bounded by things like disk transfers, though, but stuff like bandwidth to the card itself. You probably hav

A lot of deep learning workloads aren't necessarily bounded by things like disk transfers, though, but stuff like bandwidth to the card itself. You probably have enough RAM to just pre-warm your disk cache anyway if you have exceedingly large training samples. This is more useful for things like massive CAD models, if I had to guess. On Nvidia if you want this, I think you can basically achieve it using GPUDirect anyway if you really want to bypass CPU transfers and go straight from SSD to GPU, via PCIe. I guess it depends exactly on how this card is connected to the disk. Having it on-card is pretty convenient, though. GPUDirect has practical limitations on what boards/systems it will work with.