Zigqueen is a UCI chess engine written in Zig (0.15.2) with a from-scratch NNUE evaluation and a single-threaded alpha-beta search. Engine code is a clean-room implementation (see CLEAN_ROOM_RULES.md); the network is trained from the publicly published Stockfish NNUE training datasets.
Author: Stierms Rating Chess Engines Diary CEDR=3473
v6.0.0 what's new?
The full-threats NNUE release: a new evaluation network architecture and a
substantially faster inference path.
Highlights
New ZQB9 network: mirrored HalfKA with 8 king buckets, 60,144
full-threat inputs, width 1024, and eight 1024 -> 16 -> 32 -> 1
material-bucketed layer stacks.
Trained from random initialization with zigqueen's extended bullet trainer
on published Stockfish training datasets — not initialized from,
fine-tuned from, or distilled from another engine's network weights. See
NETWORK.md.
External check: 52.3% over a 1,620-game gauntlet against 27 CCRL-listed
engines (~3602 on the project's anchored gauntlet scale). Details and caveats in
STRENGTH.md.
Performance
The new architecture ships with an incremental full-threat accumulator and
nine accepted optimization rounds, together worth about +21% node throughput
on AVX2 with fixed-depth behavior preserved. The AVX2 first-layer kernel
uses a fixed accumulation order, so portable x86 builds keep cross-ISA
fixed-depth identity.
Search and time management
Time management moved to a field-confirmed 130% optimum-time spend level.
Advanced search-tuning UCI options are exposed at their tested defaults.
Platforms
Linux and Windows: x86-64-v3/AVX2 and x86-64-v4+VNNI/AVX-512 packages;
Windows builds use large pages when available, with silent fallback.
Android: armv8 and armv8-dotprod binaries, plus signed OEX APKs for
engine-GUI apps.
Compatibility
Single-threaded UCI engine; default Hash 256 MB. The embedded net is
self-contained; EvalFile can load supported external ZQB networks.
Official binaries identify as zigqueen 6.0.0.
Comments
Post a Comment