Honestly, I think we could’ve pulled off a lot earlier if GPU development had invested in GPGPU earlier.

I can see it now… the a national lab can run ImageNet, but it takes so many nodes with unobtanium 3dfx stuff that you have to wait 24 hours for a run to be scheduled and completed.