lumip

Good catch, those are typos. Parameters a and c are both negative, as you say. I'll have to fix that description, thanks!

That would indeed be a nice addition. Unfortunately I don't have any such devices at hand. I'm not even sure if their power targets can be adjusted, but even if not it would be worthwhile to include them at their reference/design power.

I'd also have to check whether the same benchmark is available for MacOS, but fusing scores from multiple benchmarks for this would be an interesting long-term goal of mine anyways.

(Thinking about the first point, I should probably highlight the design power point for each model somehow in the plot).

Thanks! That's a good question and it is, of course, just an assumption.

My main reasoning was that as we increase the power level more and more, I would expect diminishing returns, i.e., a saturation curve as we would see in many physical (and especially electromagnetical) processes. Hence concavity. I would find a relation where we could get essentially limitless compute out of the die if we just provide enough power much less likely.

I could be wrong about that, of course. It could be that the function is not (globally) concave and just has a cut-off at some upper limit (basically when the CPU blows up). Or maybe the real function is concave but doesn't follow the exponential that I used here.

In the end I settled at the current formulation somewhat out of convenience and because it seemed to be able to explain/fit the (limited) data already quite well.

I did some experimentation using concave (again!) polynomials, but found the that in that case it was much more difficult to achieve convergence - and when it did, the results weren't much different than the current ones.

Do you have any particular suggestion for an alternative model in mind? I will probably continue experimenting with this a bit more in the future.