This release makes progress primarily in FRC and short time control games. Most of the gains come from a more well-rounded network (now including ~10% DFRC data) and multiple speedups.
For the sake of my sanity (and because .NET can't/doesn't autovectorize NNUE code), this requires at least AVX2 to be performant, but should still function without it.
Picking a binary:
Start with the non-aot
binary, and pick ...-512
if your processor has Avx512.
If you want a slightly slower but more compact binary, start with ...-aot-v4
and use the first one that doesn't crash:
- v4 enables Avx512,
- v3 enables Avx256,
- v2 is a fallback for older CPU's
8+0.08:
Lizard 10.5 vs Lizard 10.4
Elo | 59.97 +- 5.14 (95%)
Conf | 8.0+0.08s Threads=1 Hash=16MB
Games | N: 4002 W: 889 L: 205 D: 2908
Penta | [3, 121, 1138, 667, 72]
http://somelizard.pythonanywhere.com/test/1141/
60+0.6:
Lizard 10.5 vs Lizard 10.4
Elo | 28.20 +- 6.79 (95%)
Conf | 60.0+0.60s Threads=1 Hash=128MB
Games | N: 1000 W: 103 L: 22 D: 875
Penta | [0, 13, 395, 90, 2]
http://somelizard.pythonanywhere.com/test/1146/
DFRC:
Lizard 10.5 vs Lizard 10.4
Elo | 108.91 +- 8.01 (95%)
Conf | 8.0+0.08s Threads=1 Hash=16MB
Games | N: 4002 W: 1760 L: 545 D: 1697
Penta | [58, 153, 640, 816, 334]
http://somelizard.pythonanywhere.com/test/1143/