Apple M2

[Modern Times]

My machine is mostly used for video editing, but it does a decent job for chess purposes.

M1 Max is a superb machine for video editing from what I've read, and those Stockfish speeds are more than good enough for 99% of people, yes.

[rjgibert]

The trouble with using Stockfish to compare the performance of an M1/M2 vs some x86 machine is that Stockfish is very heavily optimized for x86. This makes for an apple and oranges comparison. However, if you are looking for an endorsement from somebody in the know about hardware/software, it would be hard to do much better than Linus Torvalds.

[George Sobala]

Stockfish is probably as optimized as it can be for M1/2.

The problem is NNUE: this flies on x86 and does not on ARM. Classical SF is very fast on the M1 architecture.

[Modern Times]

The trouble with using Stockfish to compare the performance of an M1/M2 vs some x86 machine is that Stockfish is very heavily optimized for x86. This makes for an apple and oranges comparison.

It is a totally valid comparison if it Stockfish that you want to use. That is the real world situation.

[rjgibert]

The trouble with using Stockfish to compare the performance of an M1/M2 vs some x86 machine is that Stockfish is very heavily optimized for x86. This makes for an apple and oranges comparison.

It is a totally valid comparison if it Stockfish that you want to use. That is the real world situation.

Nevertheless, people are mistakenly concluding that the difference in performance is a hardware issue rather than software issue. So far, nobody has made that distinction clear.

[Werewolf]

Stockfish is probably as optimized as it can be for M1/2.

The problem is NNUE: this flies on x86 and does not on ARM. Classical SF is very fast on the M1 architecture.

If this is true it's very significant and appears to be a hardware - not software - issue.

[wickedpotus]

Nevertheless, people are mistakenly concluding that the difference in performance is a hardware issue rather than software issue. So far, nobody has made that distinction clear.

Who cares?
If I want to use the best engines for Chess I really dont care if it is a hardware or SW issue that makes the price/perf equation suck.

If you compile it natively and try to optimize it (even given the sucky openness and info from Apple about the low level stuff) and it still sucks for GPU and CPU-bound engines compared to similar or lower priced more open alternatives. Why go down the lock-in Apple rabbit hole?

We have yet to see someone post an actual M2 benchmark... The only post done was from Apple's +3500 USD "max" laptop (Apple currently's the fastest laptop) and wasn't even a proper bench...

On the screenshot, it showed about 9Mn/s (I tested the same position (Ne7 in Spanish on a 1k USD laptop with AMD 4800H and it was about 50% faster than the Mac showed

I agree that 9Mn/S is "ok"... But for 3500 USD *No way Jose*.. when you can get much better SF speed for a laptop 1/3 of the price and as as cream on top you can even get a good GPU on an intel or AMD laptop to run engines like Lc0 etc they way they should (and to train networks for SF)... And even do some occassional 3D gaming on the laptop


Another great thing for us chess lovers is that SF 15 has become na officlal bencmark for pts... these days...

https://openbenchmarking.org/test/pts/s ... 05#metrics

[Ras]

The trouble with using Stockfish to compare the performance of an M1/M2 vs some x86 machine is that Stockfish is very heavily optimized for x86.

Wrong. The ARM optimisations for Apple have already been in for months. The hardware simply doesn't have the same potential as x86 for the SIMD instructions that Stockfish is leveraging. Torvald's observations are entirely irrelevant here because things like compilation don't use that stuff in the first place.

[wickedpotus]

The trouble with using Stockfish to compare the performance of an M1/M2 vs some x86 machine is that Stockfish is very heavily optimized for x86.

Wrong. The ARM optimisations for Apple have already been in for months. The hardware simply doesn't have the same potential as x86 for the SIMD instructions that Stockfish is leveraging. Torvald's observations are entirely irrelevant here because things like compilation don't use that stuff in the first place.

People seem confuse the marketing stunt of releasing 5.19 kernel on linux running on an ARM-mac with Linus with switching to Mac as his workhorse..

"This time around, and on an M-series Mac, Torvalds has been running test builds and boots, however it is the Linux creator’s hope that he will one day be able to travel with, and rely upon, one of these arm64 laptops" (from techradar article)

[smatovic]

Stockfish is probably as optimized as it can be for M1/2.

The problem is NNUE: this flies on x86 and does not on ARM. Classical SF is very fast on the M1 architecture.

If this is true it's very significant and appears to be a hardware - not software - issue.

I asusme this is cos of the SIMD/vector-unit used for NNUE inference, AVX2 vs. NEON, in future we might see AVX3-512 for x86 and SVE2 for ARM based silcon.

--
Srdja