20x error in your benchmark results

[Bulat-Ziganshin]

Your work became quite famous, in particular cited in https://www.quora.com/Why-are-so-many-people-willing-to-buy-Mac-Minis-to-run-OpenClaw/answer/Franklin-Veaux

The ANE is roughly 80× more efficient per FLOP than an A100. Of course, the A100 has 50× more total throughput.

This means that A100 uses 4,000 times more energy, i.e. 13 KW. It's 20-30 times more than the real A100 power usage, and once you use real numbers, A100 is pretty close to ANE. And if you compare with a modern GPU, like 5090, any difference will be gone.

There is no Apple magic; the silicon is the same, and you can't pull 80x more efficiency than the rest of the world (which has pretty similar efficiency). You can use outdated GPUs like A100 to get 2-3x boost compared to modern silicon, but that's all.

Your work is GREAT, but unfortunately, Apple boys can't count their fingers and unconditionally believe everything that looks like Apple magic.

[usatenko]

A lot of people own macs M1,2,3,4,5 as their daily machines, and I think the excitement is because they discovered how to employ that NPU that they already have out of the box with good performance. But jumping from that to “comparing ANE and A100” is obviously a stretch.

[metacosm]

How was the power measured, actually?