This version is around 60 elo stronger than 1.10 at hyper-bullet self play.
The following pre-compiled binaries are included in the download:
- texel64-avx512.exe
For Windows 7 or later x86-64 systems using AVX-512, VNNI, POPCOUNT and BMI2 instructions. - texel64-avx2-bmi.exe:
For Windows 7 or later x86-64 systems using AVX2, POPCOUNT and BMI2 instructions. - texel64-avx2-pop.exe:
For Windows 7 or later x86-64 systems using AVX2 and POPCOUNT instructions. - texel64-avx2.exe:
For Windows 7 or later x86-64 systems using AVX2 instructions. - texel64-ssse3.exe
For Windows x86-64 systems with SSSE3 support. - texel64-old.exe
For Windows x86-64 systems. - texel64cl.exe:
Cluster version of texel64-avx2-pop.exe.
Requires Microsoft MPI to be installed. - texel-arm64-dot:
For the ARMv8-A 64-bit architecture with dot product CPU instruction.
Should work on many modern Android devices. - texel-arm64:
For the ARMv8-A 64-bit architecture.
Should work on most modern Android devices. - texel64:
For Linux x86-64 systems with AVX2 and POPCOUNT support.
There are no pre-built 32-bit binaries, but 32-bit is still supported in the source code and build system.
List of changes:
Neural network improvements:
- Fewer non-incremental updates of the first layer state.
- Implement quantization aware training.
- Better GPU utilization by using separate data loading threads.
- Efficiently handle training data that is too large to fit in RAM.
- Train a new network based on new training data evaluated by Texel 1.10.
- Optionally use AVX-512 instructions.
Other:
- Re-calibrate UCI_Elo for NNUE.
- Fix a null-move verification problem.
- Update GoogleTest code to version 1.14 to fix warnings when using new CMake versions.
- Add state to the Square class, i.e. make it contain the square number.