This release includes the following improvements and changes:
- Introduction of the new threat inputs NNUE architecture, which has since been introduced in many other top engines (including Stockfish, Reckless and others). As of late, the NNUE additionally has pawn-pair inputs
- Many speedups to make the NNUE architecture changes viable
- Greatly improved search logic, pushing the concept of fractional depth further to more heuristics
- SMP improvements: Sharing correction histories between threads and tuning under multithreaded conditions
- Proper NUMA handling on linux
- General improvements to the source code
Performance vs. PlentyChess 7.0.0
STC
Elo | 41.82 +- 2.53 (95%)
Conf | 8.0+0.08s Threads=1 Hash=16MB
Games | N: 20220 W: 6372 L: 3950 D: 9898
Penta | [37, 1488, 4824, 3538, 223]
https://furybench.com/test/6597/
LTC
Elo | 47.35 +- 2.35 (95%)
Conf | 40.0+0.40s Threads=1 Hash=64MB
Games | N: 20038 W: 6312 L: 3598 D: 10128
Penta | [2, 1229, 4912, 3805, 71]
https://furybench.com/test/6603/
DFRC
Elo | 49.46 +- 5.50 (95%)
Conf | 8.0+0.08s Threads=1 Hash=16MB
Games | N: 5064 W: 1419 L: 703 D: 2942
Penta | [28, 341, 1166, 881, 116]
https://furybench.com/test/6604/
Binary Guide
- generic: Slowest
- ssse3: Slightly faster, requires a CPU with SSSE3 support
- fma: Faster, requires a CPU with FMA support
- avx2: Faster, requires a CPU with AVX2 support
- bmi2: Faster, requires a CPU with AVX2 and BMI2 support. Not recommended for Zen1 and Zen2 CPUs
- avx512: Faster, requires a CPU with AVX512 support
- avx512vbmi2: Fastest, requires a CPU with AVX512 and VBMI2 support
- neon: For ARM CPUs that support NEON instructions
If you're unsure about what binary to use, start with avx512vbmi2 and find the first one that does not crash on your machine. There is a chance that the AVX2 binary is faster on your system, despite it supporting AVX512.