Releases: crippa1337/svart
Svart 6
I am happy to present Svart 6, featuring threaded input handling, LazySMP and other improvements to the engine.
Elo gain
Using the 8moves_v3.epd book in self-play:
8+0.08 (STC)
ELO | 20.77 +- 8.38 (95%)
CONF | 8.0+0.08s Threads=1 Hash=32MB
GAMES | N: 2512 W: 554 L: 404 D: 1554
60+0.60 (LTC)
ELO | 12.98 +- 9.36 (95%)
CONF | 60.0+0.60s Threads=1 Hash=128MB
GAMES | N: 1500 W: 241 L: 185 D: 1074
4 Thread Scaling
ELO | 120.05 +- 10.39 (95%)
CONF | 8.0+0.08s Threads=4 Hash=128MB
GAMES | N: 2512 W: 1127 L: 292 D: 1093
Noteable changes
- Node count TM (#50)
- TM formula improvement (#51)
- Asp. windows improvement (#52)
- Threaded input handling (#55)
- LazySMP (#58)
Full changelog: v5.0.0...v6.0.0
Binaries
Svart's release scheme follows the microarchitecture levels.
x86_64-v1 is the slowest but compatible with almost anything.
x86_64-v2 is faster but is compatible with CPUs pre-Haswell/Excavator.
x86_64-v3 is faster still and recommended on modern systems.
x86_64-v4 is the fastest but requires AVX-512 support.
See the readme for instructions on how to compile these versions yourself.
Svart 5
I am happy to present Svart 5, featuring a few improvements to search that had previously gone under my radar.
Elo gain
Using the Pohl.epd book in self-play:
8+0.08 (STC)
ELO | 108.61 +- 7.59 (95%)
CONF | 8.0+0.08s Threads=1 Hash=16MB
GAMES | N: 5000 W: 2254 L: 740 D: 2006
40+0.4 (LTC)
ELO | 143.15 +- 10.78 (95%)
CONF | 40.0+0.40s Threads=1 Hash=256MB
GAMES | N: 2504 W: 1232 L: 255 D: 1017
Noteable changes
- Three-fold LMR (#44)
- Internal Iterative Reductions (#46)
- Fail-Hard NMP (#48)
- Improvements to the TT scheme (#49)
Full changelog: v4.0.0...v5.0.0
Binaries
Svart's release scheme follows the microarchitecture levels.
x86_64-v1 is the slowest but compatible with almost anything.
x86_64-v2 is faster but is compatible with CPUs pre-Haswell/Excavator.
x86_64-v3 is faster still and recommended on modern systems.
x86_64-v4 is the fastest but requires AVX-512 support.
See the readme for instructions on how to compile these versions yourself.
Svart 4
I am happy to present Svart 4, featuring a much stronger evaluation network and around 180 elo gain.
Svart 4 has used an OpenBench instance to test its features.
Thanks to the contributions made by @cosmobobak, @uwuplant, @Adam-Kulju, @archishou, @Ciekce, @JouanDeag, @GediminasMasaitis, @Alex2262, @lucametehau, @spamdrew128 and @dede1751, Svart's development has been made much more efficient.
Elo gain
Using the Pohl.epd book in self-play:
8+0.08 (STC)
ELO | 185.04 +- 16.88 (95%)
CONF | 8.0+0.08s Threads=1 Hash=16MB
GAMES | N: 1504 W: 931 L: 198 D: 375
40+0.4 (LTC)
ELO | 184.14 +- 23.86 (95%)
CONF | 40.0+0.40s Threads=1 Hash=64MB
GAMES | N: 752 W: 461 L: 96 D: 195
Notable changes
- Improved network (#35 & #36)
- QWord TT Entries (#37)
- Check Extensions (#38)
- Improved RFP (#39 & #40)
- Futility Pruning (#42)
Full Changelog: v3.1.0...v4.0.0
Binaries
Svart's release scheme follows the microarchitecture levels.
x86_64-v3 is the fastest and recommended on modern systems.
x86_64-v2 is less fast but is compatible with CPUs pre-Haswell/Excavator.
x86_64-v1 is the slowest but compatible with almost anything.
See the readme for instructions on how to compile these versions yourself.
Contributors
Assets 6
Svart 3.1
This release has some large changes made to time management, namely calculating an optimum time and a maximum time for a search, which gains some 30-40 elo at both STC and LTC.
Why is this getting it's own release?
Because Svart 3, and it's previous releases, has quite a glaring issue that was made clear to me today. If Svart wouldn't complete a search of depth 1 then it would not have a best move to give and would crash. With this release, Svart is forced to always complete a search of at least depth 1 and if it can't then it will simply time out instead of crashing.
This release of Svart also introduces a new scheme of release with better optimizations from the compiler. It follows the microarchitecture levels (see README.md for instructions on how to compile it yourself).
x86_64-v3 is recommended.
x86_64-v2 works on CPUs pre Haswell/Excavator.
x86_64-v1 should work on anything.
Full Changelog: v3.0.0...v3.1.0
Svart 3
I am happy to present Svart 3, introducing an NNUE evaluation from a 786->256x2->1 net.
Thank you to:
Who's open source code and help contributed a lot to the development of Svart's NNUE.
I would also like to thank @dsekercioglu for their amazing Marlinflow trainer that Svart's nets are trained with.
Elo gain
Using the noob_4moves.epd book in self-play:
8+0.08 (STC)
Score of svart3 vs svart2 after 512 rounds: 855 - 41 - 128 (0.90)
Ptnml: WW WD DD/WL LD LL
Distr: 357 110 35 10 0
Stats: W: 83.5% L: 4.0% D: 12.5% TF: 0
Elo difference: 376.85 +/- 28.65
40+0.4 (LTC)
Score of svart3 vs svart2 after 700 rounds: 1150 - 40 - 210 (0.90)
Ptnml: WW WD DD/WL LD LL
Distr: 474 172 45 8 1
Stats: W: 82.1% L: 2.9% D: 15.0% TF: 0
Elo difference: 374.91 +/- 23.12
Noteable changes
Full changelog: v2.0.0...v3.0.0
Contributors
Assets 5
Svart 2
I am happy to present Svart 2, the second release of the Svart engine.
Elo gain
Using the noob_4moves.epd book
8+0.08
Score of svart-2 vs svart-1 after 1500 games: 1338 - 21 - 141 (0.94)
Ptnml: WW WD DD/WL LD LL
Distr: 594 130 25 1 0
Stats: W: 89.2% L: 1.4% D: 9.4%
Elo difference: 474.93 +/- 28.38
Noteable changes
- Null Move Pruning (#11)
- History Heuristic Improvements (#12) & (#15)
- Reverse Futility Pruning (#13)
- Better Mate Checks (#14)
- TT usage in QSearch (#17)
- Late Move Reductions (#18)
- TT Improvements (#19)
- Late Move Pruning (#22)
Full Changelog: v1.0.0...v2.0.0
Svart 1
Svart 1.0.0
Thank you, to everyone who has helped, teached and motivated me from the Engine Programming and Stockfish community, making this release possible.
As there is no previous release to compare to I ran 1000 games against Stash 13.0, rated 2033±22 in the CCRL.
Using the 8moves_v3 book with 8+0.08 time control:
Score of svart-1.0.0 vs stash-13.0: 473 - 391 - 136 [0.541] 1000
... svart playing White: 253 - 185 - 62 [0.568] 500
... svart playing Black: 220 - 206 - 74 [0.514] 500
... White vs Black: 459 - 405 - 136 [0.527] 1000
Elo difference: 28.6 +/- 20.1, LOS: 99.7 %, DrawRatio: 13.6 %