Stockfish NPS by Joachim26 (TriNet Engine)

Triton6654 · Post by **Triton6654** » 08/04/2024, 0:10

Experimental version by Joachim26, featuring 3 NNUE networks, source code attached.

https://pixeldrain.com/u/uNP8GniS

Triton6654 · Post by **Triton6654** » 08/04/2024, 1:26

Triton6654 wrote: ↑08/04/2024, 0:10 Experimental version by Joachim26, featuring 3 NNUE networks, source code attached.

https://pixeldrain.com/u/uNP8GniS

https://pixeldrain.com/u/LmvSsdix

Stockfish NPS did not perform well, but got a very aggressive playing style.

Crystal 8 played well despite a lower search depth n HynoS seem to be the most stable n strongest engine right now.

LittleJoe · Post by **LittleJoe** » 10/04/2024, 11:22

Triton6654 wrote: ↑08/04/2024, 0:10 Experimental version by Joachim26, featuring 3 NNUE networks, source code attached.

https://pixeldrain.com/u/uNP8GniS

There are now prereleases of SFNNv9.5.3 with 3 nets for Windows and Android available:
https://github.com/Joachim26/StockfishNPS/releases

The two main nets are gen SFNNv9 and SFNNv5, the L1=128 smallnet (no official name thus the .3 in the engine name) is the one used in Stockfish master.

For short TCs the whole game is played with the SFNNv5 net, for longer TCs more and more moves of the game are played with the new SFNNv9 net.

With "Mediumnet threshold" in the uci-option the point (move) where the medium net starts to evaluate the positions can be chosen:
0: only SFNNv9 is used in the whole game
10000: Most of the game is played with the SFNNv5 net
Default in the MOMENT: 300

mraz64 · Post by **mraz64** » 10/04/2024, 13:21

pls,ver AVX2

Post by **Hurnavich** » 10/04/2024, 14:17

Just lost a game with the latest one all 3 nets running and in folder on default not a good start

and i'm on 16 cores and 7 man TB's

[Event "Rated game, 5 min"]
[Site "Engine Room"]
[Date "2024.04.10"]
[Round "?"]
[White "Rozzlei001, HypnoS 030424 avx2"]
[Black "Squeek, SFNNv9.5.3_20240409"]
[Result "1-0"]
[ECO "C65"]
[WhiteElo "3203"]
[BlackElo "3178"]
[Annotator "0.15;0.00"]
[PlyCount "107"]
[EventDate "2024.04.10"]
[SourceTitle "playchess.com"]
[TimeControl "300"]

{HypnoS 030424 avx2 (22 threads): 33.6 plies; 21,026kN/s AMD Ryzen 9 7900X3D
12-Core Processor      10MHz, (12 cores, 24 threads), Analysis-AQ-2024.ctg,
1024 MB} 1. e4 {B 0} e5 {B 0} 2. Nf3 {B 0} Nc6 {B 0} 3. Bb5 {B 0} Nf6 {B 0} 4.
d3 {B 0} Bc5 {B 0} 5. Bxc6 {B 0} dxc6 {B 0} 6. O-O {B 0} Bg4 {B 0} 7. h3 {B 0}
Bh5 {0.00/35 29} 8. g4 {B 0 (Nbd2)} Nxg4 {0.00/32 3} 9. hxg4 {B 0} Bxg4 {
0.00/35 3} 10. Be3 {B 0} Be7 {0.00/36 4} 11. Kg2 {B 0 (Nbd2)} f5 {0.00/38 4}
12. Qe1 {B 0 (Rh1)} Bxf3+ {-0.24/28 3} 13. Kxf3 {B 0} f4 {-0.25/29 4} 14. Bd2 {
B 0} Qd6 {-0.03/36 40} 15. Bc3 {B 0} g5 {-0.03/30 6} 16. Ke2 {B 0} g4 {0.00/42
11} 17. Nd2 {B 0} Qe6 {0.00/36 1} 18. f3 {B 0 (b4)} h5 {-0.33/28 3} 19. Nc4 {
B 0 (Rh1)} Bf6 {-0.21/26 4} 20. Rh1 {B 0} O-O-O {-0.14/26 3} 21. Qf2 {B 0} Kb8
{-0.05/33 6} 22. a3 {B 0 (Qg2)} g3 {-0.23/25 3} 23. Qg2 {B 0 (Qf1)} Qe7 {
-0.04/37 11} 24. Rag1 {B 0 (Qh3)} h4 {0.00/38 7} 25. Qh3 {B 0} Qc5 {0.00/38 2}
26. Rg2 {0.15/55 29} Qe7 {0.00/45 0} 27. Kf1 {0.24/34 8 (Qf5)} c5 {0.00/39 3 
(b6)} 28. Re2 {0.53/29 6} b6 {0.31/35 16 (Bg7)} 29. Kg2 {0.35/34 17} Rh6 {
0.42/33 0 (a5)} 30. b4 {0.64/30 7} cxb4 {0.43/34 0} 31. axb4 {0.68/32 9} Kb7 {
0.64/33 11 (Bg7)} 32. Ra1 {0.80/25 6} Qe8 {0.63/32 9 (Qf8)} 33. Be1 {0.97/29 5
(Ra2)} Qf8 {0.62/25 15} 34. c3 {1.07/31 0} a6 {0.60/20 0} 35. Nb2 {1.02/33 5 
(Rea2)} Be7 {0.66/22 1} 36. Rd1 {1.05/33 3 (Nc4)} Qe8 {0.90/35 9} 37. Rc2 {
1.01/35 0} Rc6 {0.88/23 0} 38. c4 {1.00/37 4} b5 {0.88/28 1} 39. Ra1 {1.05/32 4
} Qh5 {0.90/28 0 (Qf7)} 40. Bc3 {1.27/28 4} Qf7 {0.90/30 0} 41. Rcc1 {1.73/26
5 (Rb1)} Bf6 {1.22/26 9 (Qg6)} 42. cxb5 {2.95/27 7} axb5 {1.82/24 0} 43. Ra5 {
3.13/24 4} Kb6 {2.00/26 0 (Ra6)} 44. d4 {3.71/25 5 (Ra3)} Be7 {3.10/25 7} 45.
Kg1 {4.13/26 0} Qb3 {3.27/17 1} 46. Qf1 {4.30/28 4} Rc4 {3.45/21 0} 47. Nxc4+ {
4.43/31 8} bxc4 {3.45/22 0} 48. Rxe5 {4.52/25 4 (Qd1)} Bd6 {3.74/18 2 (h3)} 49.
Ra5 {5.28/26 5 (Rh5)} h3 {4.14/27 4 (Bf8)} 50. Qxh3 {199.42/69 4 (Qd1)} Bxb4 {
4.27/14 1} 51. Qe6+ {199.59/57 3 (Qf5)} Rd6 {11.32/17 1} 52. Qc8 {#30/49 10}
Bxa5 {12.73/45 0} 53. Bxa5+ {#27/51 3} Kxa5 {199.72/48 0} 54. Qxc7+ {#24/48 5
Squeek,SFNNv9.5.3_20240409- resigns} 1-0

Triton6654 · Post by **Triton6654** » 11/04/2024, 0:51

My testing result shows that SFNNv9 240409 is on par with 11 stockfish clones which are on D3072.

XXXXX

Admin Message:
Your test results include Stockfish derivatives without source code, please read the rules again before posting them, as this violates the rule 8c, thanks.
- IbaiBuR

LittleJoe · Post by **LittleJoe** » 11/04/2024, 9:35

Please PM me your deleted results, Triton. This does not violate any forum rule I guess.
Over night I run my first(!) test with more than one thread namely with two, TC was/is 40+0.4s and the performance of SFNNv9.5.3 was/is disappointing.
I'm not sure if there is a "bug" in my code or if the reason for the bad performance comes e. g. from the tens of parameters which are not optimized for the old SFNNv5 net. To do this optimization with my hardware is just impossible.
So for the moment I suggest: If you are not satisfied with this new engine concept just set "Mediumnet threshold" to zero

or use another engine....

LittleJoe · Post by **LittleJoe** » 24/04/2024, 17:19

Could someone please inform Eduard that he just has to delete
the line:

std::cout << "OptimalTime " << SE::tmOptTime << sync_endl;

and recompile it to make Lichess happy. The line is only used during debugging, something which is useless for Eduard.
One of the SF maintainer is the origin of this Lichess/uci problem since he loves to do one refactoring after the other

Joachim will fix this after fixing much bigger problems already merged.

Post by **M.Z** » 24/04/2024, 17:44

LittleJoe wrote: ↑24/04/2024, 17:19 Could someone please inform Eduard that he just has to delete
the line:

std::cout << "OptimalTime " << SE::tmOptTime << sync_endl;

and recompile it to make Lichess happy. The line is only used during debugging, something which is useless for Eduard.
One of the SF maintainer is the origin of this Lichess/uci problem since he loves to do one refactoring after the other
Joachim will fix this after fixing much bigger problems already merged.

@LittleJoe
I had already written to him...

LittleJoe · Post by **LittleJoe** » 24/04/2024, 18:10

M.Z wrote: ↑24/04/2024, 17:44
LittleJoe wrote: ↑24/04/2024, 17:19 Could someone please inform Eduard that he just has to delete
the line:

std::cout << "OptimalTime " << SE::tmOptTime << sync_endl;

and recompile it to make Lichess happy. The line is only used during debugging, something which is useless for Eduard.
One of the SF maintainer is the origin of this Lichess/uci problem since he loves to do one refactoring after the other
Joachim will fix this after fixing much bigger problems already merged.
@LittleJoe
I had already written to him...

Thanks, if only all bugs could be solved so quickly. I already realized that there will be a potential problem when I saw Disservin's like 5th refactoring.

Today's SF updating with 3 commits including the Finny-tables will make much more problems for updating the triple net engine. And it's impossible to skip Finny's tables commit since on my PC I have just measured 15% speedup. Wow, congrats to Finny!
And thanks to Gabe and Disservin for the implementation.

Outskirts CheSS ForuM

Outskirts CheSS ForuM