Releases: Matthies/RubiChess
2.1
2.0.1 - Minor bug fix
This fixes a bug evaluating the KBNvK endgame.
You can safely stay at release 2.0 if you are using Syzygy tablebases.
2.0
Progress
20+0.2 on a BMI2+AVX machine:
Score of 2.0 vs 1.9: 295 - 165 - 740 [0.554] 1200
Elo difference: 37.79 +/- 12.11
60+0.6 on a POPCNT/SSSE3 machine:
Score of 2.0 vs 1.9: 521 - 202 - 1277 [0.580] 2000
Elo difference: 55.9 +/- 9.0
Not bad but probably far less on modern PC with longer time control and against different opponents.
1.9
This release besides some minor improvements introduces NNUE evaluation to RubiChess and with this some more binaries:
- RubiChess-BMI2: For best performance on modern intel CPUs and probably also new AMD Ryzen Zen3 / 5?00X CPUs (tests welcome)
- RubiChess-AVX2: For best performance on modern AMD Ryzen Zen/Zen2
- RubiChess: For older CPUs that support POPCNT but no AVX2
- RubiChess-SSSE3: For even older CPUs with SSSE3 but no POPCNT
- RubiChess-Legacy: For very old x86-64 CPU without SSSE3 support (are there any?)
You will get a warning at startup if the selected binary doesn't match your CPU or it will just crash.
NNUE evaluation is disabled by default. It can be enabled by UCI option "Use NNUE" and it uses default network file nn-803c91ad5c-20201107.nnue which is part of Windows package and for download at https://github.com/Matthies/NN
1.8
Three binaries now:
RubiChess-BMI2 for (Intel-)CPUs supporting the BMI2 instructions.
RubiChess for CPUs supporting POPCNT (use also on AMD CPUs)
RubiChess-Legacy for everything older
Testing 4000 games at 60+0.6/8moves_v3
Score of 1.8 vs 1.7.3: 753 - 483 - 2764 [0.534] 4000
Elo difference: 23.5 +/- 6.0, LOS: 100.0 %, DrawRatio: 69.1 %