f4bcd40409
Update nnue architecture to SFNNv16 and net nn-89cb98a217f7.nnue
...
Failed STC (https://tests.stockfishchess.org/tests/view/6a4f36355529b8472df7ff7e ):
LLR: -2.94 (-2.94,2.94) <0.00,2.00>
Total: 144768 W: 38003 L: 38051 D: 68714
Ptnml(0-2): 521, 17154, 37050, 17170, 489
Passed LTC (https://tests.stockfishchess.org/tests/view/6a501dad5529b8472df80103 ):
LLR: 2.94 (-2.94,2.94) <0.50,2.50>
Total: 189888 W: 49613 L: 48984 D: 91291
Ptnml(0-2): 122, 20406, 53256, 21041, 119
Passed VLTC (https://tests.stockfishchess.org/tests/view/6a52892c5529b8472df80481 ):
LLR: 2.94 (-2.94,2.94) <0.00,2.00>
Total: 46124 W: 12156 L: 11865 D: 22103
Ptnml(0-2): 12, 4523, 13696, 4824, 7
Passed VVLTC (https://tests.stockfishchess.org/tests/view/6a566a285529b8472df80b84 ):
LLR: 2.94 (-2.94,2.94) <0.50,2.50>
Total: 44698 W: 11805 L: 11507 D: 21386
Ptnml(0-2): 9, 3819, 14395, 4117, 9
This PR adds pawn pair features, a new set of features that is indexed by pairs of pawns occupying the same or adjacent files (hence "3-wide"). Because these are a superset of pawn–pawn interactions in threat inputs (including the recently added pawn-pusher inputs), those are removed. HalfKA features remain unchanged.
Pawn-pair features were invented by Jonathan for Pawnocchio, based on his observation that in a net trained on *all* pairs of pawns, the most important pawn pairs were those differing by at most one file.
sscg13 made the necessary changes to the trainer (https://github.com/official-stockfish/nnue-pytorch/pull/502 ) and initial changes to inference. The slowdown was originally far too large (some 10%); following major optimizations from Jonathan and a bit from me, the slowdown is small enough for a comfortable LTC pass; on my machine I get a 3.5% slowdown overall. I expect there are further "easy" speedups in the new code, so this gap should narrow a bit.
Nettest PR: https://github.com/vondele/nettest/pull/359
closes https://github.com/official-stockfish/Stockfish/pull/6982
Bench: 2718396
Co-Authored-By: anematode <timothy.herchen@gmail.com >
Co-Authored-By: Jonathan Hallström <lmj.hallstrom@gmail.com >
2026-07-20 11:06:41 +02:00
sscg13 and Joost VandeVondele
de7b0cd7a6
Update trace eval
...
Updates the information printed in the "eval" command. Removes piece value estimation, and adds both small and big net eval information.
closes https://github.com/official-stockfish/Stockfish/pull/6743
No functional change
2026-04-19 07:10:20 +02:00
sscg13 and Joost VandeVondele
a6d055d7e2
Update NNUE architecture to SFNNv13 and net nn-5227780996d3.nnue
...
Since the introduction of Threat Inputs in SFNNv10 and its refinements, the
accumulator/L1 size has significantly decreased. This has made it more
computationally favorable to double the L2 size from 16 to 32.
Failed STC:
LLR: -2.98 (-2.94,2.94) <0.00,2.00>
Total: 189184 W: 49649 L: 49635 D: 89900
Ptnml(0-2): 881, 22479, 47812, 22585, 835
https://tests.stockfishchess.org/tests/view/69928ef7bfe62dc95a624d5c
Passed LTC:
LLR: 2.94 (-2.94,2.94) <0.50,2.50>
Total: 37902 W: 9833 L: 9509 D: 18560
Ptnml(0-2): 35, 4074, 10408, 4400, 34
https://tests.stockfishchess.org/tests/view/69937936ff782f8eeef611e3
Passed VLTC SMP:
LLR: 2.94 (-2.94,2.94) <0.50,2.50>
Total: 20576 W: 5455 L: 5191 D: 9930
Ptnml(0-2): 6, 1837, 6338, 2101, 6
https://tests.stockfishchess.org/tests/view/6994df0586e4a32314116399
closes https://github.com/official-stockfish/Stockfish/pull/6619
bench: 2396996
2026-02-18 21:46:57 +01:00
sscg13 and Joost VandeVondele
83e42045a6
Update NNUE architecture to SFNNv12 and net nn-7e1657811c6d.nnue
...
This net is trained with https://github.com/sscg13/nettest/blob/no-king-attacker/threats.yaml .
Passed non-regression STC:
LLR: 2.93 (-2.94,2.94) <-1.75,0.25>
Total: 92640 W: 24371 L: 24209 D: 44060
Ptnml(0-2): 389, 11047, 23322, 11137, 425
https://tests.stockfishchess.org/tests/view/6986345fa259e26d2c32e9c5
Passed non-regression LTC:
LLR: 2.96 (-2.94,2.94) <-1.75,0.25>
Total: 198636 W: 50960 L: 50919 D: 96757
Ptnml(0-2): 176, 22218, 54511, 22215, 198
https://tests.stockfishchess.org/tests/view/6987ee29b0f3ca5200aafa51
closes https://github.com/official-stockfish/Stockfish/pull/6609
bench: 2465828
2026-02-13 06:56:10 +01:00
sscg13 and Joost VandeVondele
fac506bdf3
Update NNUE architecture to SFNNv11 and net nn-3dd094f3dfcf.nnue
...
Removes threat features of the form piece -> king, thus saving 13MB of net
space and approximately 0.8 threat feature updates per incremental accumulator
update.
Passed non-regression STC:
LLR: 2.93 (-2.94,2.94) <-1.75,0.25>
Total: 33664 W: 8864 L: 8636 D: 16164
Ptnml(0-2): 136, 3926, 8501, 4112, 157
https://tests.stockfishchess.org/tests/view/6981dcda4776a4e6e7fef2ac
Passed non-regression LTC:
LLR: 2.94 (-2.94,2.94) <-1.75,0.25>
Total: 46896 W: 12077 L: 11881 D: 22938
Ptnml(0-2): 41, 5118, 12924, 5334, 31
https://tests.stockfishchess.org/tests/view/69827c6beb87369ff4d0c7d5
closes https://github.com/official-stockfish/Stockfish/pull/6593
bench: 2668754
2026-02-04 18:26:28 +01:00
sscg13 and Disservin
32292d1e62
Represent threat weights directly as i8
...
LEB128 adds no additional compression and adds extra complexity to the
code currently.
closes https://github.com/official-stockfish/Stockfish/pull/6479
No functional change
2025-12-21 15:43:32 +01:00
+1
8e5392d79a
Update NNUE architecture to SFNNv10 with Threat Inputs and net nn-49c1193b131c.nnue
...
This commit introduces Full Threat Input features, which are a subset of Piece(Square)-Piece(Square) pairs. In any given position, the active features consist of pairs where the second piece’s square lies in the attack set of the first piece. This is an extremely simplified explanation that leaves out many details. The already-used HalfKAv2_hm feature set completes the input features.
Minor quantization changes have also been made.
The net nn-49c1193b131c.nnue was trained by vondele using the following setup: https://github.com/vondele/nettest/blob/7de71238e9b295e3f88ed7c9c5936af632c9b981/threats.yaml
A graphical version of an earlier scheme (with less refinement) that illustrates the core concepts can be found attached.
[NewInputs.pdf](https://github.com/user-attachments/files/23478441/NewInputs.pdf )
Further information, as well as a brief description of the history of development, can be found attached.
[Stockfish threat inputs PR summary.pdf](https://github.com/user-attachments/files/23478634/Stockfish.threat.inputs.PR.summary.pdf )
This has been a huge effort spanning over half a year, with the original [discussion thread](https://discord.com/channels/435943710472011776/1336647760388034610 ) reaching over 11k messages. Thanks to everyone who has contributed.
Monty PRs:
https://github.com/official-monty/Monty/pull/87 (Initial threat input PR)
https://github.com/official-monty/Monty/pull/114 (Fixed threat indexing to take into account colour correctly)
https://github.com/official-monty/Monty/pull/116 (i8 quantisation of weights whilst keeping calculations in i16)
Yukari commit:
https://github.com/yukarichess/yukari/commit/2d482c64a79cec03cf4987d5289334b9cdc737bc (Threat inputs merged)
Plentychess PRs:
https://github.com/Yoshie2000/PlentyChess/pull/400 (Threat inputs merged)
https://github.com/Yoshie2000/PlentyChess/pull/411 (Threat input weights quantised to i8)
Passed STC:
LLR: 2.93 (-2.94,2.94) <0.00,2.00>
Total: 63424 W: 16956 L: 16591 D: 29877
Ptnml(0-2): 276, 7522, 15797, 7795, 322
https://tests.stockfishchess.org/tests/view/69105b3dec1d00d2c195c569
Passed LTC:
LLR: 2.95 (-2.94,2.94) <0.50,2.50>
Total: 27876 W: 7417 L: 7110 D: 13349
Ptnml(0-2): 23, 3033, 7530, 3318, 34
https://tests.stockfishchess.org/tests/view/6910d817ec1d00d2c195c66e
Passed VVLTC (Hash accidentally set to 1/2 normal value for both sides):
LLR: 2.94 (-2.94,2.94) <0.50,2.50>
Total: 12458 W: 3353 L: 3102 D: 6003
Ptnml(0-2): 0, 1106, 3767, 1355, 1
https://tests.stockfishchess.org/tests/view/69115a26ec1d00d2c195c7cd
This version has also passed non-regression LTC against the originally passed version:
LLR: 2.94 (-2.94,2.94) <-1.75,0.25>
Total: 51144 W: 13086 L: 12903 D: 25155
Ptnml(0-2): 22, 5167, 15018, 5336, 29
https://tests.stockfishchess.org/tests/view/69138a317ca87818523314bf
LTC elo estimate on ARM:
1 patch : 13.9 1.9 38296.5 73728 52
2 master : 0.0 ---- 35431.5 73728 48
closes https://github.com/official-stockfish/Stockfish/pull/6411
bench: 2626086
Co-authored-by: Shawn Xu <xu107288696@gmail.com >
Co-authored-by: Timothy Herchen <timothy.herchen@gmail.com >
Co-authored-by: Viren6 <94880762+Viren6@users.noreply.github.com >
Co-authored-by: Yoshie2000 <patrick.leonhardt@gmx.net >
Co-authored-by: Joost Vandevondele <Joost.VandeVondele@gmail.com >
Co-authored-by: rn5f107s2 <clemens.lerchl@gmail.com >
Co-authored-by: cj5716 <125858804+cj5716@users.noreply.github.com >
Co-authored-by: AliceRoselia <63040919+AliceRoselia@users.noreply.github.com >
Co-authored-by: Linmiao Xu <linmiao.xu@gmail.com >
Co-authored-by: Disservin <disservin.social@gmail.com >
2025-11-12 10:49:39 +01:00
sscg13 and Disservin
7858f9dfdc
Use same pawn value in both nets when doing material scaling of eval
...
Passed Non-regression STC:
LLR: 2.94 (-2.94,2.94) <-1.75,0.25>
Total: 221312 W: 57291 L: 57274 D: 106747
Ptnml(0-2): 760, 26152, 56841, 26117, 786
https://tests.stockfishchess.org/tests/view/676e2a101a2f267f20548453
Passed Non-regression LTC:
LLR: 2.94 (-2.94,2.94) <-1.75,0.25>
Total: 176808 W: 45084 L: 45023 D: 86701
Ptnml(0-2): 112, 19418, 49286, 19473, 115
https://tests.stockfishchess.org/tests/view/676f424d1a2f267f2054857f
closes https://github.com/official-stockfish/Stockfish/pull/5741
Bench: 1121800
2025-01-12 21:09:58 +01:00
