It's impressive they got it to work — the lowest I'd heard of this far was native FP8 training.