a bug in stockfish?

[Uri Blass]

[d]6r1/1k6/p7/Pp1p4/1PpP4/2P2PP1/3P2K1/8 w - - 0 1

stockfish is only at depth 12 after some minutes

FEN: 6r1/1k6/p7/Pp1p4/1PpP4/2P2PP1/3P2K1/8 w - - 0 1

Stockfish_25052105_x64_modern:
Available processors: 0-7
Using 6 threads
NNUE evaluation using nn-1c0000000000.nnue (133MiB, (22528, 3072, 15, 32, 1))
NNUE evaluation using nn-37f18f62d772.nnue (6MiB, (22528, 128, 15, 32, 1))
1/3 00:00 64 64k -1.22 g3-g4
2/3 00:00 142 142k -1.08 g3-g4
3/4 00:00 887 887k -1.10 g3-g4 Rg8-g6 Kg2-g1
4/5 00:00 946 946k -1.10 g3-g4 Rg8-g6 Kg2-g1 Rg6-h6
5/6 00:00 3k 1,414k -1.41 g3-g4 Kb7-c8 Kg2-g1
6/8 00:00 5k 2,265k -1.66 g3-g4 Kb7-c8 Kg2-f1 Rg8-f8 Kf1-f2 Rf8-f7
7/8 00:00 8k 2,721k -1.96 g3-g4 Kb7-c8 Kg2-f2 Rg8-f8 Kf2-g3 Kc8-c7 Kg3-f2
8/10 00:00 11k 2,767k -1.79 f3-f4 Rg8-e8 Kg2-f2 Kb7-c8
9/18 00:00 45k 2,970k -2.29 Kg2-f2 Kb7-c8 Kf2-g2 Kc8-d8 f3-f4 Kd8-e7
10/27 00:00 76k 2,607k -2.28 Kg2-f2 Kb7-c8 Kf2-g2 Kc8-d8 Kg2-f2 Kd8-e8 g3-g4 Rg8-f8 Kf2-g1 Ke8-e7
11/35 00:00 249k 3,079k -2.11 Kg2-f2 Kb7-c8 Kf2-g2 Kc8-d8 Kg2-f2 Rg8-f8 Kf2-g2 Kd8-c8 Kg2-f2 Rf8-e8 Kf2-f1 Re8-h8
12/16+ 00:10 90,033k 8,702k -2.02 Kg2-f2
12/18+ 02:18 1,294,912k 9,335k -1.85 Kg2-f2

[kasinp]

Thanks for posting this Uri!

Indeed Stockfish seems stumped by this. But also, Crystal 9 seems not to recognize that White is doing just fine.
An immediate 0.00 eval from Crystal 8 though!

Analysis by Crystal 8:

1.Kf2 Kc8 2.Kg2 Kc7 3.Kf2
= (-0.03) Depth: 8/64 00:00:01
= (0.00) Depth: 43/19 00:00:08 414MN, tb=89451
1.Kf2 Kc8 2.Kg2 Kc7 3.Kf2
= (0.00) Depth: 44/19 00:00:09 443MN, tb=108674
1.Kf2 Kc8 2.Kg2 Kc7 3.Kf2
= (0.00) Depth: 45/19 00:00:09 475MN, tb=120346
1.Kf2 Kc8 2.Kg2 Kc7 3.Kf2
= (0.00) Depth: 46/19 00:00:10 504MN, tb=132556
1.Kf2 Kc8 2.Kg2 Kc7 3.Kf2
= (0.00) Depth: 47/19 00:00:11 543MN, tb=166118
1.Kf2 Kc8 2.Kg2 Kc7 3.Kf2
= (0.00) Depth: 48/19 00:00:12 600MN, tb=201992
1.Kf2 Kc8 2.Kg2 Kc7 3.Kf2
= (0.00) Depth: 49/19 00:00:13 640MN, tb=226451
1.Kf2 Kc8 2.Kg2 Kc7 3.Kf2
= (0.00) Depth: 50/19 00:00:13 654MN, tb=239838
1.Kf2 Kc8 2.Kg2 Kc7 3.Kf2
= (0.00) Depth: 51/19 00:00:14 699MN, tb=268840

Peter

[tmokonen]

How odd. The more threads you use, the slower the search in this position.

[Uri Blass]

downloaded latest stockfish version with 6 cores.

Still horrible branching factor and no more data after more than 40 minutes with 6 cores

FEN: 6r1/1k6/p7/Pp1p4/1PpP4/2P2PP1/3P2K1/8 w - - 0 1

Stockfish_25052306_x64_avx2:
Available processors: 0-7
Using 6 threads
NNUE evaluation using nn-1c0000000000.nnue (133MiB, (22528, 3072, 15, 32, 1))
NNUE evaluation using nn-37f18f62d772.nnue (6MiB, (22528, 128, 15, 32, 1))
1/3 00:00 63 63k -1.22 g3-g4
2/3 00:00 125 125k -1.08 g3-g4
3/4 00:00 1k 1,300k -1.14 g3-g4 Kb7-a8
4/6 00:00 4k 1,757k -1.42 g3-g4 Kb7-c8
5/6 00:00 4k 1,950k -1.52 g3-g4 Kb7-c8 Kg2-f1 Rg8-g6 Kf1-g1
6/7 00:00 4k 2,182k -1.43 g3-g4 Kb7-c8 Kg2-f1 Rg8-g6 Kf1-g1 Rg6-f6
7/9 00:00 6k 3,142k -1.61 g3-g4 Kb7-c8 Kg2-h1 Rg8-g6
8/12 00:00 13k 4,250k -1.82 g3-g4 Rg8-f8 Kg2-g3 Kb7-c7 Kg3-f2 Kc7-d8
9/24 00:00 23k 4,511k -1.91 Kg2-h2 Kb7-c8 Kh2-g2 Rg8-f8
10/20 00:00 53k 4,440k -2.16 Kg2-h2 Rg8-e8 Kh2-g1 Re8-e2 g3-g4 Kb7-c8 g4-g5 Re2-e8 f3-f4 Kc8-d7
11/25 00:00 104k 3,714k -2.41 Kg2-f2 Kb7-c8 Kf2-g2 Kc8-d7 Kg2-f2 Kd7-e6 Kf2-g2 Ke6-f6 Kg2-f2 Kf6-f5 Kf2-g2 Rg8-g6 Kg2-f2
12/22 00:10 104,039k 10,150k -2.24 Kg2-f2 Kb7-c8 Kf2-g2 Kc8-d7 Kg2-f2 Kd7-e6 Kf2-g2 Ke6-f5 Kg2-f2 Rg8-e8 Kf2-f1 Kf5-g5 Kf1-f2 Re8-e7 Kf2-f1 Kg5-f5 Kf1-f2 Re7-e6 Kf2-f1
13/21+ 05:56 3,433,495k 9,638k -2.09 Kg2-f2

[Uri Blass]

tried with a single core and nothing better(bigger depth but still horrible branching factor at depth 18 and no data at depth 19 after more than 40 minutes.

FEN: 6r1/1k6/p7/Pp1p4/1PpP4/2P2PP1/3P2K1/8 w - - 0 1

Stockfish_25052306_x64_avx2:
Available processors: 0-7
Using 1 thread
NNUE evaluation using nn-1c0000000000.nnue (133MiB, (22528, 3072, 15, 32, 1))
NNUE evaluation using nn-37f18f62d772.nnue (6MiB, (22528, 128, 15, 32, 1))
1/3 00:00 10 10k -1.22 g3-g4
2/3 00:00 22 22k -1.08 g3-g4
3/4 00:00 99 99k -1.04 g3-g4 Rg8-g6
4/6 00:00 663 332k -1.57 g3-g4 Rg8-e8 Kg2-g1 Kb7-c8
5/6 00:00 703 352k -1.55 g3-g4 Rg8-e8 Kg2-g1 Kb7-c8 Kg1-f1
6/7 00:00 893 447k -1.60 g3-g4 Kb7-c8 Kg2-g1 Rg8-g6
7/8 00:00 3k 443k -1.91 Kg2-h2 Kb7-c8 g3-g4 Rg8-g6 Kh2-g3 Rg6-f6
8/9 00:00 3k 397k -1.84 g3-g4 Kb7-c8 Kg2-f2 Rg8-g6 Kf2-g3 Kc8-d7 Kg3-h2 Rg6-f6
9/17 00:00 4k 483k -2.04 g3-g4 Kb7-c8 Kg2-f2 Rg8-g6 Kf2-g3 Rg6-h6 f3-f4 Rh6-h1 Kg3-g2 Rh1-e1 Kg2-f2
10/18 00:00 6k 570k -2.00 g3-g4 Kb7-c8 Kg2-f2 Rg8-g6 Kf2-g3 Kc8-d7 Kg3-g2 Rg6-f6 Kg2-g3 Rf6-f7 Kg3-g2 Kd7-e7 f3-f4 Rf7xf4
11/17 00:00 9k 578k -1.77 g3-g4 Rg8-e8 Kg2-g1 Kb7-c8 Kg1-f2
12/16 00:00 15k 594k -2.19 Kg2-f2 Kb7-c8 Kf2-g2 Rg8-e8 Kg2-f1 Kc8-d8 f3-f4 Kd8-e7 Kf1-f2 Ke7-f6
13/21 00:00 21k 693k -2.29 f3-f4 Kb7-c8 Kg2-f3 Kc8-d7 g3-g4 Rg8-e8 f4-f5 Kd7-e7 g4-g5 Re8-g8
14/19 00:00 30k 715k -2.42 Kg2-f2 Kb7-c8 Kf2-g2 Kc8-d7 Kg2-f2 Kd7-e7 Kf2-g2 Ke7-f6 Kg2-f2 Kf6-g5 Kf2-g1 Rg8-e8
15/22 00:00 31k 726k -2.41 Kg2-f2 Kb7-c8 Kf2-g2 Kc8-d7 Kg2-f2 Kd7-e7 Kf2-g2 Ke7-f6 Kg2-f2 Kf6-g5 Kf2-g1 Rg8-e8 Kg1-f1 Kg5-f5 Kf1-f2 Kf5-f6
16/33 00:00 50k 807k -2.29 Kg2-f2 Kb7-b8 Kf2-g2 Kb8-c8 Kg2-f2 Rg8-e8 Kf2-f1 Kc8-d7 Kf1-f2 Re8-g8 Kf2-g2 Kd7-d8 Kg2-f2 Rg8-e8 g3-g4 Kd8-e7
17/36 00:00 61k 830k -2.28 Kg2-f2 Kb7-b8 Kf2-g2 Kb8-c8 Kg2-f2 Rg8-e8 Kf2-f1 Kc8-d7 Kf1-f2 Re8-g8 Kf2-g2 Kd7-d8 Kg2-f2 Kd8-e7 Kf2-g2 Ke7-f6 Kg2-f2 Rg8-e8 Kf2-f1 Kf6-f5 Kf1-f2 Re8-e7 Kf2-f1 Re7-e6 Kf1-f2 Kf5-f6 Kf2-f1 Kf6-f7 Kf1-f2 Re6-e8
18/52 00:03 5,896k 1,823k -2.07 Kg2-f2 Kb7-b8 Kf2-g2 Kb8-c8 Kg2-f2 Rg8-e8 Kf2-f1 Kc8-b7 Kf1-f2 Kb7-b8 Kf2-f1 Kb8-c8 Kf1-f2 Kc8-d8 Kf2-f1 Kd8-d7 Kf1-f2 Re8-e6 Kf2-f1 Re6-e7 Kf1-f2 Kd7-e8 Kf2-f1 Re7-e6 Kf1-f2 Re6-h6 Kf2-f1

[jefk]

which GUI ?
it may be a gui glitch (or not)
with Obsidian it (still) gets stuck at about 25 ply in the (old i admit) chpartner interface
but runs fine in old uncle fritz (19)

[Jouni]

Weird position. SF says after long search hashfull 0! But Obsidian behaves similarly.

[peter]

Dont' have latest SF dev. till now, but with 20250513 it's similiar as for analysis- output and time to depth. Crystal 9 at my trial gets stuck very soon too and also with for Black winning eval, which is plausible as for material only of course. Lc0 also sees Black in advantage but "only" -/+, so tried to let it play out with White against SF. Net 6147500PT, 3070ti GPU, 8G NNcache, SF with 28 threads of 16x3x5GHz CPU, 8G hash, 25'+5" TC:
[pgn][Event "AMD, Schnellschach 25.0min+5.0sek"]
[Site "Dort"]
[Date "2025.05.25"]
[Round "1"]
[White "Lc0 v0.32.0-dev+git.63b66e4"]
[Black "Stockfish"]
[Result "1/2-1/2"]
[Annotator "-1.00;-2.76"]
[SetUp "1"]
[FEN "6r1/1k6/p7/Pp1p4/1PpP4/2P2PP1/3P2K1/8 w - - 0 1"]
[PlyCount "97"]
[GameId "2181487681625843"]
[TimeControl "1500+5"]

1. Kf2 {-1.00/13 43} Kc8 {-2.76/12 357} 2. Kg2 {-0.98/11 2} Kd7 {-2.50/8 306 (Kb8)} 3. Kf2 {-0.98/14 34} Re8 {-2.32/8 222 (Kc8)} 4. Kf1 {-0.97/15 46} Re7 {-2.04/9 168 (Kc8)} 5. Kf2 {-0.92/13 7} Rg7 {-1.96/10 175} 6. Kg2 {-0.98/13 0} Re7 {-1.90/9 142 (Kc7)} 7. Kf1 {-0.96/16 136 (Kf2)} Re6 {-2.23/9 109 (Kc8)} 8. Kf2 {-0.93/12 5} Kc7 {-1.80/8 44 (Tg6)} 9. Kf1 {-0.97/15 37} Re8 {-1.73/8 0 (Kb8)} 10. Kf2 {-0.98/13 35} Kb8 {-1.67/9 0 (Tg8)} 11. Kf1 {-1.00/13 33} Re6 {-1.58/10 0 (Th8)} 12. Kf2 {-0.98/15 33} Kc7 {-1.90/8 0 (Te8)} 13. Kf1 {-0.97/15 33} Kd8 {-1.65/8 1 (Kb8)} 14. Kf2 {-0.94/17 34} Rg6 {-1.57/8 4} 15. Kg2 {-0.96/16 17} Kc7 {-1.70/8 21 (Kc8)} 16. Kf2 {-0.98/15 35} Kc8 {-1.82/8 0 (Kb8)} 17. Kg2 {-0.98/15 40} Rh6 {-1.46/8 0 (Kb8)} 18. Kg1 {-0.97/15 33} Rh7 {-1.49/10 0 (Tg6)} 19. Kg2 {-0.97/15 36} Rg7 {-1.50/8 1} 20. Kf2 {-0.98/15 31} Rg6 {-1.44/8 10 (Kb8)} 21. Kg2 {-0.98/15 35} Kc7 {-1.46/7 1 (Kb8)} 22. Kf2 {-0.98/15 34} Rf6 {-1.32/9 6 (Tg8)} 23. Kg2 {-0.97/15 37} Kd7 {-1.67/9 1 (Tf8)} 24. Kf2 {-0.95/15 32} Rg6 {-1.35/10 10 (Kc8)} 25. Kg2 {-0.96/16 33} Kd8 {-1.50/9 5 (Kc8)} 26. Kf2 {-0.95/15 34} Rf6 {-1.32/8 4 (Tg5)} 27. Kg2 {-0.95/16 34} Kd7 {-1.24/8 4 (Tf5)} 28. Kf2 {-0.95/15 31} Kc8 {-1.76/8 8} 29. Kg2 {-0.97/15 21} Rh6 {-1.26/8 15 (Tf8)} 30. Kg1 {-0.97/15 33} Rh8 {-1.17/8 0 (Kb8)} 31. Kg2 {-0.98/15 27} Rg8 {-1.54/8 5} 32. Kf2 {-0.99/14 27} Re8 {-1.36/9 5 (Kb8)} 33. Kf1 {-0.97/14 39} Kd8 {-1.22/8 0 (Tg8)} 34. Kf2 {-0.95/14 39} Kd7 {-1.24/10 0 (Kc8)} 35. Kf1 {-0.96/14 35} Re7 {-1.08/8 5 (Kc7)} 36. Kf2 {-0.94/15 37} Re6 {-1.02/10 3 (Tg7)} 37. Kf1 {-0.93/16 31} Ke8 {-1.02/10 11 (Kc7)} 38. Kf2 {-0.88/16 31} Kf7 {-1.14/7 6 (Kd7)} 39. Kf1 {-0.81/15 31} Kg8 {-1.10/10 4 (Te8)} 40. Kf2 {-0.76/16 31} Kh8 {-1.05/9 5 (Kf7)} 41. Kf1 {-0.62/14 30} Kg8 {-0.87/9 6} 42. Kf2 {-0.64/13 2} Rf6 {-0.61/9 34 (Kf7)} 43. Kg2 {-0.62/12 40} Kf8 {-0.09/13 0} 44. Kf2 {-0.64/12 29} Kg8 {0.00/25 0 (Tc6)} 45. Kg2 {-0.52/11 33} Rg6 {0.00/44 0 (Tc6)} 46. Kf2 {-0.13/10 30} Kg7 {0.00/83 0 (Td6)} 47. Kg2 {-0.04/8 26} Rf6 {0.00/92 0 (Th6)} 48. Kf2 {-0.02/6 25} Rg6 {0.00/79 7 (Te6)} 49. Kg2 {0.00/4 24 Remis angenommen} 1/2-1/2
[/pgn]
So at least as for playing out in this one pairing and hardware- TC, SF finds out draw being best for Black to get against Lc0 with White, I could imagine in infinite analysis its fortress detection does work well enough in starting position to "see" the lack of progress for Black but not well enough to get 0.00- near eval with reasonable time to depth (time to eval in this case).

Lc0- output at test position:

6r1/1k6/p7/Pp1p4/1PpP4/2P2PP1/3P2K1/8 w - - 0 1

Analysis by Lc0 v0.32.0-dev+git.63b66e4:

1.Kf2 Ka8 2.Kg2 Th8 3.Kg1 Tb8 4.Kf1 Ka7 5.Kf2 Kb7 6.Kg1 Td8 7.Kf1 Th8 8.Kg1 Th6 9.Kg2 Te6 10.Kf2 Kc6 11.Kf1
Tiefe: 15/30 00:03:51 43818kN
-/+ (-1.00)
And after some more ponder- time:
1.Kf2 Ka7 2.Kg2 Th8 3.Kg1 Tb8 4.Kf1 Ka8 5.Kf2 Kb7 6.Kg1 Ta8 7.Kg2 Th8 8.Kg1 Kc6 9.Kg2 Te8 10.Kf2 Kd6 11.Kf1
Tiefe: 16/32 00:15:46 123MN
Schwarz steht besser.
-/+ (-1.00)

So as for time to depth (if it can be compared to A-B at all) there isn't so much difference to SF here anyhow, yet eval is nearer to the truth (considering the big difference as for material that's misleading of course), and even if -1.00 still yet also is about as near to 0.00, which was the only really correct one, as it is to the -+ of SF's, so not very decisive neither. Just knowing, that with this one material's advantage anything beneath clear winning eval of Black's means draw, just from this one practcially experienced point of view, Lc0 yet evaluates much better here

[royb]

I'm pretty sure it is not related to the GUI used. Hiarcs Chess Explorer Pro on a Mac shows Stockfish doing the same thing. I tried the latest Stockfish development version from the abrok website. Same issue. I tried Stockfish 17.1 - same issue. Stockfish 17 - same issue.

Stockfish 16 does not exhibit this strange behavior - it makes progress just fine in its search. Seems likely a bug introduced sometime after Stockfish 16.

[peter]

Indeed Stockfish seems stumped by this. But also, Crystal 9 seems not to recognize that White is doing just fine.
An immediate 0.00 eval from Crystal 8 though!

Yes, Crystal 8 shines here.
Yet not only Crystal 9 doubt's the fortress, Crystal 5WK and Crystal 6PM didn't come to 0.00- near eval quickly neither at my trials. Crystallite 1.1a, user dorsz posted in OpenChess once way back then derived from Crystal5, get's it (0.00- eval) almost at once too.
As for the idea, it could depend on GUI, as for SF dev. 20250513, which is the latest one I downloaded so far, it's about the same as for analysis- output in console as it is in GUI, regards