I’d be interested to hear how long it takes, wall clock, to train the same model to the same loss with same number of batches. I don’t trust utilization to say

I’d be interested to hear how long it takes, wall clock, to train the same model to the same loss with same number of batches. I don’t trust utilization to say anything useful.