Reminder: Stockfish is best with 2 cores

[Jouni]

Yes in testsuites more cores don't help. I tried HTC110 and 1 minute limit. Stockfish 17.1 scores 76 with 6 cores and 78 with 2 cores.

[Werewolf]

I've tested this extensively and see a considerable jump from 1 core to 2 cores, but then everything after that is basically noise with no gain going from 2 cores to 64 cores.

However, it's important to stress this is a tactical testsuite, not Elo.

[Werewolf]

Weirdly, the Stockfish Cluster even on low core counts (like 64 cores) scores well above normal Stockfish and continues to improve as the hardware gets bigger going all the way to 2000 cores +

I have not heard a good explanation of why this is.

[gordonr]

https://www.talkchess.com/forum/viewtop ... 9&start=60

[gordonr]

I've tested this extensively and see a considerable jump from 1 core to 2 cores, but then everything after that is basically noise with no gain going from 2 cores to 64 cores.

However, it's important to stress this is a tactical testsuite, not Elo.

Provide example positions. You don't need to share your whole test suite. Just enough positions that I can do many test runs with and see the average times.

[Werewolf]

I've tested this extensively and see a considerable jump from 1 core to 2 cores, but then everything after that is basically noise with no gain going from 2 cores to 64 cores.

However, it's important to stress this is a tactical testsuite, not Elo.

Provide example positions. You don't need to share your whole test suite. Just enough positions that I can do many test runs with and see the average times.

I'm not sharing any from my testsuite for obvious reasons - it takes ages to go through a public testsuite and remove cooks / errors etc.

Since this phenomenon is well attested I don't see the point anyway. Even the SF team kindof admitted it on Discord.

However, if I can find tough positions that are sound that I'm not using for testing - I'll share them.

[Jouni]

Eduard Nemeth's REMS is really stunning solver. With correct parameters it solved 102/110 with 6 cores and 1 minute. Very difficult to beat!

[Ciekce]

in testsuites

problem found and conclusion discarded

[Uri Blass]

No evidence that 2 cores is best for stockfish in test suites or even that 2 cores are better than 6 cores.

Even if there is one test suite that stockfish solve more positions with 2 cores relative to 6 cores it does not prove the conclusion but we even did not see one test suite so people can test to see if they can reproduce the results.

Stockfish is not deterministic with more than one core so 76 vs 78 may be the opposite in another run.

[Uri Blass]

I can add that I am not sure what is HTC110 that is mentioned in the first post

searching in google leaded me to

https://open-chess.org/viewtopic.php?t=4026 that has different number than 110.