Nyxptr - this release focuses on internal architecture changes to improve search efficiency and fix performance issues on Intel CPUs.
Author: František Šilhán
⚡ Key Improvements
NN Batching: The engine now groups 16 positions together before sending them to the neural network. This significantly increases Nodes Per Second (NPS) by utilizing CPU SIMD instructions more effectively.
Intel Performance Fix: Resolved a major bug where Intel CPUs experienced massive slowdowns (90% drop) due to thread overhead and denormal floating-point numbers.
Memory Efficiency: Added pre-allocated memory buffers and direct bitboard-to-tensor conversion, reducing CPU time spent on memory management.
Consistent Search: Stabilized NPS across different hardware configurations by restricting LibTorch to single-thread inference per request.
📦 Assets
Nyxptr-v0.0.6-windows-cpu.zip: The complete standalone package for Windows. Includes the engine binary, all required DLLs, and the v4 model file.
Nyxptr-Model-v4.zip: The neural network weights provided separately for Linux users and source builders.
Comments
Post a Comment