To be fair, I think ML workloads are quite a bit different than the days of compiling over lunch breaks.
What the above post was probably trying to get at is that the ML specific hardware is far more efficient these days than consumer GPUs.
What the above post was probably trying to get at is that the ML specific hardware is far more efficient these days than consumer GPUs.
Ah yes, my code can't be useful to people unless it takes a long time to compile...