I am pleased to release Spaghet BologNNese 1.0, which is up 150 selfplay elo stronger depending on the opening book.
I've implemented:
- Output buckets (8 linear buckets)
- SIMD Inference, which is about a 20% speedup
- Lazy SMP multithreading
Estimated CCRL Elo: ~3500
Features to look forward to in the next release:
- Thread Voting
- King buckets
- Quadratic output buckets
Progression Tests
STC Pohl:
Elo | 151.82 +- 10.56 (95%)
Conf | 8.0+0.08s Threads=1 Hash=8MB
Games | N: 1508 W: 762 L: 142 D: 604
Penta | [0, 16, 215, 410, 113]
https://chess.aronpetkovski.com/test/2143/
LTC Pohl:
Elo | 147.44 +- 10.90 (95%)
Conf | 40.0+0.40s Threads=1 Hash=64MB
Games | N: 1006 W: 476 L: 73 D: 457
Penta | [0, 9, 122, 332, 40]
STC 8moves:
Elo | 139.12 +- 10.45 (95%)
Conf | 8.0+0.08s Threads=1 Hash=8MB
Games | N: 1504 W: 609 L: 37 D: 858
Penta | [0, 13, 263, 367, 109]
LTC 8moves:
Elo | 109.83 +- 10.96 (95%)
Conf | 40.0+0.40s Threads=1 Hash=64MB
Games | N: 1000 W: 322 L: 16 D: 662
Penta | [0, 7, 218, 237, 38]
SPECIAL THANKS:
for sharing an instance, hardware, and ideas,
- Zuppa
- fireandice
- Ciekce
- Vast
- Somelizard
for helping me with NNUE and other search features,
and @noobpwnftw for generously donating hardware to furybench.