Reminder: Stockfish is best with 2 cores

[Werewolf]

retesting...give me a few days...

[Jouni]

Probably my 100€ CPU is not NUMA . BTW I have 2 additional testsuites about 100 positions. Very similar behaviour: Stockfish don't improve practically after 2 cores. Stockfish 16.1 had one clever patch causing this "problem".

[gordonr]

Probably my 100€ CPU is not NUMA . BTW I have 2 additional testsuites about 100 positions. Very similar behaviour: Stockfish don't improve practically after 2 cores. Stockfish 16.1 had one clever patch causing this "problem".

You know what I'm going to say Are you ok to post the positions here? How many runs did you average over? I presume you're saying this behaviour applies to the latest Stockfish dev versions too? Did you mean 16.1 or 17.1?

[Werewolf]

Still testing…btw my comments apply to SF17 and 17.1 only. I’ve not tested the latest dev versions but will pick up the progress when SF18 comes out (hopefully soon, it has been 6 months).

IIRC someone worked out this odd behaviour began with 16.1.

@ Jouni - ugh i thought I’d found the explanation for a moment.

[Werewolf]

Also have you gone through the linked suite properly yet and are you claiming there's nothing there?

See the last post from the previous discussion. Yes, I spent time looking at the linked test sets and they were full of cooks; contrived positions; etc. Hence I switched to just asking "show me the actual FENs that support the claim". Nobody has been able to do that. Any positions I've tried have not supported the claim.

I noticed on your video that 1 core can solve 112/162 with an average of about 7 seconds. So 69% of your tests are pretty basic if they are solved that quick on 1 core. I realise you had a 60 second limit but just how basic are your tests? I could come up with a test set that is so basic that any amount of cores will find it easy enough to be equal. 7 seconds with one core is not a difficult tactical test. I think your testing is flawed and that's why you can't explain the results.

Again, until you actually post some FENs, your claim remains nonsense.

Testing now 2/3 through. I am seeing slightly better results for higher core counts, but it's very small. What I'm doing differently this time is time-logging every position.
My testing method is always to thoroughly examine every legal move of every test position, in Aquarium first.

As to the actual testing/timing method I use Fritz and "process test set". As long as one is super careful - and I am - that all the setting match like: Hash size (clear hash first), no opening book, threads, GUI tablebases = off, engine tablebases = on (pathway correctly set and confirmed), no background programs running etc etc. There's not much to go wrong. I even got ChatGPT to look over the log file to check there's no issues.

On 2 cores after 3 runs average = 110, on 16 cores 114. So slightly better for 16 cores with a slightly lower solve time. Got my hands on a 96 core machine this afternoon and it got 120/162. It's not much to shout about.

Anyway, you've repeatedly asked for positions. Are you claiming the Arasan is completely useless? Because I'm not using it and I doubt all the positions are bad. That would immediately give a point for discussion. While I'm finishing up testing if you could source some reliable positions, we can test. I'll best the suite has at least half a dozen viable positions.

[Werewolf]

Here's one from Arasan that gets repeated a bit. I have not checked this for cooks though.

[d]R4bk1/2Bbp2p/2p2pp1/1rPp4/3P4/4P2P/4BPPK/1q1Q4 w - - 0 1

2 cores:
4.98, 4.23, 12.83

16 cores:
5.02, 3.55, 8.01

[gordonr]

Here's one from Arasan that gets repeated a bit. I have not checked this for cooks though.

[d]R4bk1/2Bbp2p/2p2pp1/1rPp4/3P4/4P2P/4BPPK/1q1Q4 w - - 0 1

2 cores:
4.98, 4.23, 12.83

16 cores:
5.02, 3.55, 8.01

Looks like many cooks so not suitable. Also the average solve time is very quick - if a position is too easy, it's no surprise that lesser hardware can be almost as good... I could have a test suite of "mate in 1" positions and show that more threads don't solve quicker

Code: Select all

35/71	07:47	 6,382,882k	13,645k	+2.52	Ra7 Be8 g4 Qb2 Ra8 Kf7 Bg3 Rb3 g5 Bg7 h4 Ra3 Rb8 Qa1 Qxa1 Rxa1 Rb7 Kf8 Bg4 f5 Be2 Ra2 Bd1 Ra1 Bf3 h6 Kg2 hxg5 hxg5 Ra2 Rc7 Ra1 Be2 Ra2 Rb7
 35/91	07:47	 6,382,882k	13,645k	+2.95	Qd2 Rb4 Bd3 Qb2 Qe1 Kg7 Bf4 e5 Bg3 Kf7 Ra7 Rb7 Qa5 Be7 dxe5 fxe5 Rxb7 Qxb7 Bxe5 Kg8 Be2 Qc8 Qa7 Kf7 Kg1 Qd8 Bd4 Bf5 g4 Bd7 Kg2 Kf8 Qb7 Be8
 35/75	07:47	 6,382,882k	13,645k	+3.68	Rd8 Qxd1 Bxd1 Rb1 Bg4 Bxg4 hxg4 Kf7 Bg3 Ra1 Rc8 Ra6 Rb8 Ra4 Rb6 Ra8 Rxc6 Ke8 Bf4 Kd8 Rb6 Ke8 Bg3 Bh6 Rb7 Bf8 Bf4 h6 Rb6 Kd8 Bg3 Ke8 Kh3 Bg7 c6 Rc8 Bb8 g5 Kg3 e5 e4 dxe4
 35/75	07:47	 6,382,882k	13,645k	+4.14	Bf4 g5 Bg3 Kg7 Ra7 Be8 h4 gxh4 Bxh4 Qe4 Bxf6+ Kxf6 Bxb5 cxb5 Ra8 Kf7 Qh5+ Qg6 Qxd5+ e6 Qf3+ Qf5 Ra7+ Kf6 Qxf5+ exf5 Ra6+ Ke7 Kg3 Kd7 Kf4 Bg6 Rb6 Bh6+ Kg3 Bg5 Rxb5 Kc8 d5 h6 c6 Be7 f3 Bd6+ Kf2 Bf7 e4 fxe4
 35/73	07:47	 6,382,882k	13,645k	+5.58	Qa4 Bxh3 Bxb5 cxb5 Qa2 Qxa2 Rxa2 Bd7 Ra6 Be8 c6 h5 Ba5 e6 Kg1 Bd6 c7 Bxc7 Bxc7 Kf7 Rb6 g5 Bd8 g4 f3 gxf3 gxf3 h4 Kf2 Kg7 Rxe6 Bg6 Rb6 Bc2
 

[Werewolf]

Here's one from Arasan that gets repeated a bit. I have not checked this for cooks though.

[d]R4bk1/2Bbp2p/2p2pp1/1rPp4/3P4/4P2P/4BPPK/1q1Q4 w - - 0 1

2 cores:
4.98, 4.23, 12.83

16 cores:
5.02, 3.55, 8.01

Looks like many cooks so not suitable. Also the average solve time is very quick - if a position is too easy, it's no surprise that lesser hardware can be almost as good... I could have a test suite of "mate in 1" positions and show that more threads don't solve quicker

Code: Select all

35/71	07:47	 6,382,882k	13,645k	+2.52	Ra7 Be8 g4 Qb2 Ra8 Kf7 Bg3 Rb3 g5 Bg7 h4 Ra3 Rb8 Qa1 Qxa1 Rxa1 Rb7 Kf8 Bg4 f5 Be2 Ra2 Bd1 Ra1 Bf3 h6 Kg2 hxg5 hxg5 Ra2 Rc7 Ra1 Be2 Ra2 Rb7
 35/91	07:47	 6,382,882k	13,645k	+2.95	Qd2 Rb4 Bd3 Qb2 Qe1 Kg7 Bf4 e5 Bg3 Kf7 Ra7 Rb7 Qa5 Be7 dxe5 fxe5 Rxb7 Qxb7 Bxe5 Kg8 Be2 Qc8 Qa7 Kf7 Kg1 Qd8 Bd4 Bf5 g4 Bd7 Kg2 Kf8 Qb7 Be8
 35/75	07:47	 6,382,882k	13,645k	+3.68	Rd8 Qxd1 Bxd1 Rb1 Bg4 Bxg4 hxg4 Kf7 Bg3 Ra1 Rc8 Ra6 Rb8 Ra4 Rb6 Ra8 Rxc6 Ke8 Bf4 Kd8 Rb6 Ke8 Bg3 Bh6 Rb7 Bf8 Bf4 h6 Rb6 Kd8 Bg3 Ke8 Kh3 Bg7 c6 Rc8 Bb8 g5 Kg3 e5 e4 dxe4
 35/75	07:47	 6,382,882k	13,645k	+4.14	Bf4 g5 Bg3 Kg7 Ra7 Be8 h4 gxh4 Bxh4 Qe4 Bxf6+ Kxf6 Bxb5 cxb5 Ra8 Kf7 Qh5+ Qg6 Qxd5+ e6 Qf3+ Qf5 Ra7+ Kf6 Qxf5+ exf5 Ra6+ Ke7 Kg3 Kd7 Kf4 Bg6 Rb6 Bh6+ Kg3 Bg5 Rxb5 Kc8 d5 h6 c6 Be7 f3 Bd6+ Kf2 Bf7 e4 fxe4
 35/73	07:47	 6,382,882k	13,645k	+5.58	Qa4 Bxh3 Bxb5 cxb5 Qa2 Qxa2 Rxa2 Bd7 Ra6 Be8 c6 h5 Ba5 e6 Kg1 Bd6 c7 Bxc7 Bxc7 Kf7 Rb6 g5 Bd8 g4 f3 gxf3 gxf3 h4 Kf2 Kg7 Rxe6 Bg6 Rb6 Bc2
 

I confirm it is cooked.
However, “fast solving” doesn’t equal “mate in 1.” Mate in 1 would flatten out all hardware, but positions that solve in a few seconds are still useful because Fritz can measure even to a fraction of a second. Throwing them out wastes information.
Going from 3 seconds to 6 is a doubling of time, as is going from 3 minutes to 6 minutes (I much prefer the latter, but we’ll have very few to choose from if that’s the criteria).

I'm now going through Arasan and am shocked at what I'm seeing, cooks galore. But I'll keep searching.

[Werewolf]

looks sound

[d]1r4k1/4pp1p/pp1pq1p1/r2R4/PbP1P3/1P1QBP1P/R5P1/7K w - - 0 1

averaged over 3 runs
2 cores: 11.08s
16 cores: 12.95s

[gordonr]

looks sound

[d]1r4k1/4pp1p/pp1pq1p1/r2R4/PbP1P3/1P1QBP1P/R5P1/7K w - - 0 1

averaged over 3 runs
2 cores: 11.08s
16 cores: 12.95s

Code: Select all

SF dev 24th Aug 2025, AMD Ryzen 9 5950X 16-Core Processor (3.40 GHz), 2 GB hash, 6 men TBs

100 runs

2 threads:

13.9, 7.49, 13.02, 9.74, 13.15, 10.9, 16.3, 8.24, 4.72, 14.38, 11.25, 6.3, 36.06, 9.19, 6.98, 15.07, 1.35, 7.33, 9.92, 25.73, 
9.14, 52.39, 7.56, 14.18, 19.37, 10.94, 6.28, 15.05, 15.05, 14.73, 10.93, 9.95, 11.27, 18.47, 12.06, 9.19, 18.56, 12.14, 12.1, 3.91, 
9.07, 14.55, 19.2, 18.16, 22.59, 16.56, 9.38, 2.74, 8.99, 12.64, 14.05, 11.25, 11.66, 10.17, 8.98, 1.23, 3.2, 5.24, 3.42, 40.82, 
14.43, 5.34, 13.96, 6.61, 32.61, 9.16, 12.07, 3.4, 13.02, 4.88, 9.79, 18.28, 14.75, 11.83, 19.5, 20.37, 4.78, 14.25, 16.32, 11.81, 
6.89, 12.15, 12.3, 8.44, 6.55, 22.23, 6.38, 8.77, 7.48, 16.05, 10.59, 12.1, 1.46, 11.24, 9.47, 10.25, 7.28, 11.2, 6.74, 12.84, 

Mean: 12.26 secs

[1.23, 1.35, 1.46, 2.74, 3.2, 3.4, 3.42, 3.91, 4.72, 4.78, 4.88, 5.24, 5.34, 6.28, 6.3, 6.38, 6.55, 6.61, 6.74, 6.89, 6.98, 7.28, 7.33, 7.48, 7.49, 7.56, 8.24, 8.44, 8.77, 8.98, 8.99, 9.07, 9.14, 9.16, 9.19, 9.19, 9.38, 9.47, 9.74, 9.79, 9.92, 9.95, 10.17, 10.25, 10.59, 10.9, 10.93, 10.94, 11.2, 11.24, 11.25, 11.25, 11.27, 11.66, 11.81, 11.83, 12.06, 12.07, 12.1, 12.1, 12.14, 12.15, 12.3, 12.64, 12.84, 13.02, 13.02, 13.15, 13.9, 13.96, 14.05, 14.18, 14.25, 14.38, 14.43, 14.55, 14.73, 14.75, 15.05, 15.05, 15.07, 16.05, 16.3, 16.32, 16.56, 18.16, 18.28, 18.47, 18.56, 19.2, 19.37, 19.5, 20.37, 22.23, 22.59, 25.73, 32.61, 36.06, 40.82, 52.39]

Median: 11.25 secs

------

16 threads:

5.33, 3.66, 2.2, 4.0, 4.11, 5.4, 1.32, 4.42, 4.18, 6.73, 3.91, 2.67, 4.21, 5.68, 7.21, 2.19, 3.83, 3.05, 4.98, 2.85, 
5.0, 7.28, 6.72, 3.82, 3.16, 8.15, 9.25, 4.32, 2.2, 4.9, 6.79, 5.45, 12.95, 7.07, 5.16, 4.82, 10.44, 13.05, 7.04, 5.4, 
19.18, 10.26, 5.84, 6.11, 7.1, 2.01, 9.8, 7.7, 13.13, 7.63, 11.34, 7.62, 12.5, 3.77, 5.08, 9.39, 8.19, 2.03, 4.26, 7.84, 
5.46, 2.62, 4.92, 5.5, 9.5, 13.8, 7.02, 9.22, 7.21, 7.24, 6.69, 4.19, 10.51, 3.92, 7.17, 11.49, 7.09, 1.26, 4.19, 11.34, 
5.58, 12.32, 6.81, 3.94, 9.6, 7.67, 5.34, 6.67, 7.66, 11.98, 7.17, 5.45, 9.39, 9.81, 5.07, 13.94, 11.88, 13.38, 6.13, 3.97, 

Mean: 6.8 secs

[1.26, 1.32, 2.01, 2.03, 2.19, 2.2, 2.2, 2.62, 2.67, 2.85, 3.05, 3.16, 3.66, 3.77, 3.82, 3.83, 3.91, 3.92, 3.94, 3.97, 4.0, 4.11, 4.18, 4.19, 4.19, 4.21, 4.26, 4.32, 4.42, 4.82, 4.9, 4.92, 4.98, 5.0, 5.07, 5.08, 5.16, 5.33, 5.34, 5.4, 5.4, 5.45, 5.45, 5.46, 5.5, 5.58, 5.68, 5.84, 6.11, 6.13, 6.67, 6.69, 6.72, 6.73, 6.79, 6.81, 7.02, 7.04, 7.07, 7.09, 7.1, 7.17, 7.17, 7.21, 7.21, 7.24, 7.28, 7.62, 7.63, 7.66, 7.67, 7.7, 7.84, 8.15, 8.19, 9.22, 9.25, 9.39, 9.39, 9.5, 9.6, 9.8, 9.81, 10.26, 10.44, 10.51, 11.34, 11.34, 11.49, 11.88, 11.98, 12.32, 12.5, 12.95, 13.05, 13.13, 13.38, 13.8, 13.94, 19.18]

Median: 6.67 secs