Another thing that's underrated is system-level power efficiency (i.e., including power for CPU, networking, etc.). After all, data center power draw is more than just the cards, you also need to consider the entire system
and the data center efficiency (PUE) itself.
We can calculate the
. Not pictured is the H100 SuperPOD, but the answer to that is
.
A proper calculation also factors in PUE, which is how many excess watts the data center draws to generate one watt. For instance, a PUE of 1.1 is excellent, which means that for every 1 watt used to generate compute, 1.1 watts are pulled from the grid. H100 and A100 systems generally have a PUE of 1.25, while Atlas 900/Ascend 910 is rated at the much better 1.15.
Finally, I'll point out that energy efficiency gains aren't purely tied to node scaling. NVIDIA's B200 and H100 are both using TSMC 4N, and yet the energy efficiency of the B200 is ~65% better than the H100. At a card level (i.e., not system level as calculated above) B200 generates 2.3 FP16 TFLOPS/W vs. H100's 1.4 FP16 TFLOPS/W. So we can't write off further efficiency improvements for Huawei's next accelerator either, even if they stick with SMIC's N+2.