Yep. Same thing I found a while ago and updated (though I charted it but didn’t graph at the time). The CB24 wall data is pretty good Imho, not ideal but directionally good because we’re capturing enough information about the CPU through the chain (including PMIC stuff). The only gripe I have is that FP is not that important, but the thing is, it’s still modestly to strongly correlated with Integer performance and efficiency, so it’s directionally again a fine indicator.
And yeah one thing you’ll find is the QC stuff is like an intermediate point between Apple and X86 stuff generally. With the next one it should be not quite Apple but within the 10-25% type stuff on perf at a given wattage in ST or MT*, like on phones now
*MT is always easy to game though, like as you see with Intel/AMD one thing they love doing is comparing an MT result at the midpoint/optimal range of a bigger core’s efficiency and with *more* of them than Apple and comparing to an M chip at the peak of it’s range. Regardless it is true their disadvantage is less pronounced in MT though even if you tried to “equalize” and just compare a platform with similar total big cores throughout their curves and this is *partially* because of all the overhead that I’ve talked about here, which Lunar Lake fixed on power delivery & with cache but still only got them to about Qualcomm Gen 1 screwup tier efficiency —their architecture is just worse ofc.
The Arrow Lake result is funny because it says what I think most here would have for years now. Intel’s architecture is the big problem, not the fabs. The fabs make it all worse obviously, but like fabs (ok, broadly speaking within some variation) aren’t actually why Intel or AMD’s idle and low load power can’t compete with Apple, or why their IPC is lower & the performance curves in ST are so horizontal vs Apple (and Arm’s cores).
