3 ms·

I'm guessing H100 has 2x host energy overhead for connecting those GPUs? That might offset some of the perf/W benefits of nvidia's offering.