Stockfish 20251130 rocks!

[Jouni]

testing vs SF 17:

Code: Select all

Elo: 42.24 ± 4.0 (95%) LOS: 100.0%
Total: 4902 W: 1605 L: 1012 D: 2285
Ptnml(0-2): 0, 183, 1495, 770, 3
nElo: 102.67 ± 10.1 (95%) PairsRatio: 4.22
60+0.6 th 8 

There is also fix for SF getting stuck during search (issue #5023). I haven't tested yet.

[Damas Clásicas]

testing vs SF 17:

Code: Select all

Elo: 42.24 ± 4.0 (95%) LOS: 100.0%
Total: 4902 W: 1605 L: 1012 D: 2285
Ptnml(0-2): 0, 183, 1495, 770, 3
nElo: 102.67 ± 10.1 (95%) PairsRatio: 4.22
60+0.6 th 8 

There is also fix for SF getting stuck during search (issue #5023). I haven't tested yet.

The Elo gains achieved by Stockfish over the past month have been truly impressive! I can’t remember the SF team delivering such a significant improvement in a long time.

The previous regression test showed +32 Elo (1 core) and +29 Elo (8 cores). And now we’re seeing more than +10 Elo on top of that in both tests:

Elo: 44.76 ± 1.3 (95%) — LOS: 100.0%
Elo: 40.45 ± 1.3 (95%) — LOS: 100.0%

Massive THANKS to all SF contributors!

[Vinvin]

testing vs SF 17:

Code: Select all

Elo: 42.24 ± 4.0 (95%) LOS: 100.0%
Total: 4902 W: 1605 L: 1012 D: 2285
Ptnml(0-2): 0, 183, 1495, 770, 3
nElo: 102.67 ± 10.1 (95%) PairsRatio: 4.22
60+0.6 th 8 

There is also fix for SF getting stuck during search (issue #5023). I haven't tested yet.

The Elo gains achieved by Stockfish over the past month have been truly impressive! I can’t remember the SF team delivering such a significant improvement in a long time.

The previous regression test showed +32 Elo (1 core) and +29 Elo (8 cores). And now we’re seeing more than +10 Elo on top of that in both tests:

Elo: 44.76 ± 1.3 (95%) — LOS: 100.0%
Elo: 40.45 ± 1.3 (95%) — LOS: 100.0%

Massive THANKS to all SF contributors!

And it seems it still +5 to +10 Elo in the pipe (Testing Queue).
We will have a great SF 18 for Christmas

[Jouni]

There was 9 months without real progress. But where does this sudden +10 come? Nobody knows. SF development is pure trial and error .

[pohl4711]

There was 9 months without real progress. But where does this sudden +10 come? Nobody knows.

I know:
It came from Vebatim-net patch (just a trick, making engines running faster, when a bunch of instances of one engine is running simultaneously (singlethreaded) on one machine).
And from the new Threat-Input neural-nets: This was a substancial change, which is now optimized with several patches after indtroducing it. So, some more (small!) Celo gain will perhaps follow, but after this, the Stockfish progress will stall again, I am afraid.

[Jouni]

Verbatim patch didn't gain anything in NCM with 128-thread EPYC 7702!

[pohl4711]

Verbatim patch didn't gain anything in NCM with 128-thread EPYC 7702!

Yes, because they do not test singlethread-mode.

"Each server plays 16 games concurrently with 30+0.3 time controls. Hash is set to 128MB, and Threads is set to 8."

[Sylwy]

........ Celo gain will perhaps follow, but after this, the Stockfish progress will stall again, I am afraid.

Well, for a while YES....but there are enough theoretical paths to progress....to integrate all controlled squares and king safety in the network.....8D Problem Solving Matrix......and I could list for a long time. The team developing Stockfish has the necessary expertise....maybe (maybe) others will come.....overcoming the wall that has been reached sounds challenging for some....