I personally would stick to FP32 across the board for my ML efforts, but we have an entire cottage industry of people coming up with approximations to drive up

I personally would stick to FP32 across the board for my ML efforts, but we have an entire cottage industry of people coming up with approximations to drive up perf and perf/W, all of which will prove irrelevant until Moore's Law runs out IMO. And even then, I'll still stick to FP32 personally. Speaking from direct experience, bulletproof mixed precision is tough.