Stockfish big net and tuning

[Jouni]

Testing 11.4. version against SF16.

Code: Select all

Elo: 23.53 ± 1.4 (95%) LOS: 100.0%
Total: 60000 W: 17489 L: 13431 D: 29080
Ptnml(0-2): 75, 5095, 15696, 8965, 169
nElo: 47.82 ± 2.8 (95%) PairsRatio: 1.77

60000 @ 60+0.6 th 1

Code: Select all

Elo: 35.21 ± 1.4 (95%) LOS: 100.0%
Total: 45320 W: 14123 L: 9546 D: 21651
Ptnml(0-2): 10, 2484, 13116, 7019, 31
nElo: 79.89 ± 3.3 (95%) PairsRatio: 2.83

60000 @ 60+0.6 th 8

1 core is regression, but who cares. Is there 1 core prosessors anymore ?

[Modern Times]

Looks like the SF team are now focused on results on big hardware like TCEC and CCC uses. Although that also has relevance for home users now who often have hardware marketed as 20+ threads

[Uri Blass]

Testing 11.4. version against SF16.

Code: Select all

Elo: 23.53 ± 1.4 (95%) LOS: 100.0%
Total: 60000 W: 17489 L: 13431 D: 29080
Ptnml(0-2): 75, 5095, 15696, 8965, 169
nElo: 47.82 ± 2.8 (95%) PairsRatio: 1.77

60000 @ 60+0.6 th 1

Code: Select all

Elo: 35.21 ± 1.4 (95%) LOS: 100.0%
Total: 45320 W: 14123 L: 9546 D: 21651
Ptnml(0-2): 10, 2484, 13116, 7019, 31
nElo: 79.89 ± 3.3 (95%) PairsRatio: 2.83

60000 @ 60+0.6 th 8

1 core is regression, but who cares. Is there 1 core prosessors anymore ?

I have more than 1 core but I use one core for testing because in this way I can play some games at the same time on the same computer so 1 core is clearly useful.

[Viz]

1 core regression only at bullet TC.
There were literally no changes to SMP code in a long time and 60+0.6 8 threads is used as "very very long time control" because using 480+4.8 on 1 thread instead will make a lot of machines run out of RAM.
So you can more or less say that gains at 60+0.6 8 threads should be equal to gains on smth like 360+3.6 single thread.
Also SF team is focusing on the biggest TC it can realistically test, for all tests it's absolutely not feasible to test at 60+0.6 8 threads.
But there are some exceptions - namely VVLTC search tunings, especially ones that make singular extensions stronger (which is known to scale non-linearly for some years at this point) and also bigger L1 size of the net, since it's also known to bring more at longer time controls, albeit not always.

[Modern Times]

Makes sense.