Wall clock speed does not correlate linear to power consumption (in the general case). To oversimplify, consider decoding singlethreaded on one core versus multithreaded on 8 or 20 cores.

In some special cases multiplying runtime might be enough to achieve some accuracy, but that’s not the whole story.

It would be interesting to see this experiment done so that the decode is done enough times for each format so the power consumption differences could actually be measured on a power/watt meter.