After a four-month hiatus, version 1.25 of the Caissa chess engine has been released. Currently ranked 9th on the CEDR leaderboard, does this new version pose a real threat to Stockfish?
Caissa - UCI chess engine, Author: Michaล Witanowski
Rating Chess Engines Diary CEDR=3709
Caissa 1.25 - what's new?
This release brings a new GPU-accelerated training pipeline, a much larger neural network with 32 king buckets, NUMA support, numerous search refinements and critical bug fixes.
Changes
Evaluation
Expanded network to 32 king buckets (~48 MB, up from ~17 MB), trained for over 348 billion iterations. Four new nets were trained this cycle.
Simplified endgame evaluation, removing redundant special cases.
Search
Shared correction history tables across threads (+6 Elo on 4 threads).
Run singular search as non-cut-node.
Triple singular extensions.
Secondary continuation correction history.
Update correction history when best move is a bad capture by SEE.
Prior counter-move history update.
Adjusted fail-high score in multi-cut.
Disabled "deeper" extensions at root node.
Changed QSearch move count pruning logic.
Eagerly update accumulator in PV nodes and from depth 3 instead of 1 (~2% speedup).
Move Ordering
Improved capture scoring.
Eliminated floating-point operations from move scoring to avoid FP issues across platforms and compilers.
SMP & NUMA
NUMA support: worker threads are pinned to NUMA nodes with local memory allocation, reducing cross-node latency on multi-socket systems.
Second round of SMP parameter tuning at LTC (+3.1 Elo on 4 threads).
Performance
Optimized move generation and threat computation codegen (e.g. 26% fewer instructions in ComputeThreats).
Replaced phaddd/vphaddd with shuffle+add in NNUE horizontal-add helpers.
Optimized accumulator copy during accumulator refresh.
Excluded own king from occupied squares in threat computation, correctly modeling sliding attacks through the king's square.
Bug Fixes
Fixed accumulator update overflow at high ply depths by cancelling matching add/remove feature pairs inline during collection.
Fixed stack overflow crash at deep plies by extracting probcut into a NO_INLINE method.
Fixed time losses on ultra-short time controls.
Fixed incorrect node counting.
Many other minor bug fixes.
Infrastructure
New CUDA-accelerated trainer for GPU-based neural network training (optional, requires CUDA 11.0+ and an NVIDIA Pascal or newer GPU).
Centralized version definition.
Debug symbols in makefile for meaningful crash dumps.
Fixed macOS build.
Added more unit tests.
Progression test
STC (8+0.08)
Elo | 19.99 +- 2.58 (95%)
Conf | 8.0+0.08s Threads=1 Hash=8MB
Games | N: 19396 W: 5312 L: 4197 D: 9887
Penta | [65, 1871, 4766, 2876, 120]
SMP LTC (40+0.4, 4 threads)
Elo | 20.94 +- 3.28 (95%)
Conf | 40.0+0.40s Threads=4 Hash=256MB
Games | N: 10000 W: 2551 L: 1949 D: 5500
Penta | [0, 871, 2667, 1451, 11]
Caissa 1.24 JA vs other engines:
| Obsidian 16.15 SE | 59.5/141 | -22 | 141 games |
| Clover 9.1 JA | 68.5/135 | +2 | 135 games |
| Quanticade 3.0 | 52/97 | +7 | 97 games |
| Integral 7.0.0 JA | 40.5/87 | -6 | 87 games |
| Berserk 20250622 | 42/85 | -1 | 85 games |
| Reckless 0.9.0 dev-0dd5b9ac | 37/75 | -1 | 75 games |
| Viridithas 19.0.1 | 38.5/74 | +3 | 74 games |
| Dragon 3.3 | 35/70 | +0 | 70 games |
| Pawnocchio 1.9.2 | 34/70 | -2 | 70 games |
| Stormphrax 7.0.0 JA | 34.5/69 | +0 | 69 games |
| Alexandria 8.1.2 JA | 34/69 | -1 | 69 games |
| Motor 0.9.0 | 34/60 | +8 | 60 games |
| Cinder 0.4.1 | 30.5/57 | +4 | 57 games |
| PlentyChess Alamat 7.0.41 SE | 24/50 | -2 | 50 games |
| PZChessBot 6.0 JA | 34.5/48 | +21 | 48 games |
| Horsie 1.1.0 | 24/44 | +4 | 44 games |
| Reckless 0.9.0 dev-2a847427 | 21.5/44 | -1 | 44 games |
| Integral 7 dev.8bab JA | 16.5/36 | -3 | 36 games |
| Berserk 20250606 | 17/34 | +0 | 34 games |
| Alexandria 8.1.12 | 15.5/30 | +1 | 30 games |
| Horsie 1.1.0 JA | 15/28 | +2 | 28 games |
| PlentyChess 7.0.37 | 9.5/27 | -8 | 27 games |
| Tarnished v5.0 (Hero) | 14/26 | +2 | 26 games |
| Pawnocchio 1.9 | 13/26 | +0 | 26 games |
| Tarnished 5.0 | 13/26 | +0 | 26 games |
| Halogen 16.0.0 | 11.5/26 | -3 | 26 games |
| Vafra 12.8 Cfish | 16/24 | +8 | 24 games |
| CorChess 20251130 | 9.5/24 | -5 | 24 games |
| Stockfish 17.1 | 9/22 | -4 | 22 games |
| Viridithas 18.0.0 | 11.5/21 | +2 | 21 games |
| Bread Engine 3.0.0 | 15/20 | +10 | 20 games |
Comments
Post a Comment