2 ms·

If training requires N operations, using more GPUs (that otherwise would idle) simply means you finish (and iterate) faster.