Code 0.9.4
New Network
- Switched the training dataset to BT4-scored LC0 training data.
- First hidden layer after FT (what we call L1) switched from 32-wide to 16-wide with dual-activation.
- Added an L1 penalty on the feature-transformer's output activations, a regulariser that significantly sharpens eval quality.
- Various other changes to the training recipe.
Search Improvements
- Margin-aware LMP -- late-move pruning gains a predictive margin dimension
- Time Management - stop allocating less time when the eval is rising.
- Increased MAX_PLY to increase the depth cap.
- Added scaling of eval as our internal scoring to the externally reported score.
- Fixed various points where we could report a truncated PV.
- After an extensive set of ablation tests, removed a range of search and move-gen features that were now net-negative.
Performance
- Better pre-fetching support in the feature transformer and in correction-history.
- Skip eager threat-delta generation in low-piece searchers.
- Added fused kernel support for AVX2 at L1=16.
- Explicit hugetlb tier for the big weight matrices.
Strength Gain
+38.3 vs 0.9.3 at STC (https://ob.atwiss.com/test/3274/)
Elo | 38.27 +- 1.14 (95%)
Conf | 10.0+0.10s Threads=1 Hash=64MB
Games | N: 100006 W: 30860 L: 19890 D: 49256
Penta | [279, 7766, 23378, 17866, 714]
+36.5 vs 0.9.3 at LTC (https://ob.atwiss.com/test/3275/)
Elo | 36.49 +- 2.12 (95%)
Conf | 40.0+0.40s Threads=1 Hash=256MB
Games | N: 25000 W: 7471 L: 4855 D: 12674
Penta | [16, 1828, 6226, 4384, 46]
Internal cross-engine testing in a pool of the top engines at various TC shows 15-25 Elo gains over 0.9.3.
Which binary?
Which x86-64 binary? Take v3 on anything from roughly 2013 onwards (Haswell or newer / any Ryzen) — it's measurably faster (~12% NPS) than v2 thanks to the newer compiler baseline. Take v2 only if v3 exits with an illegal-instruction error on your older hardware. NNUE SIMD kernels (AVX2/AVX-512-VNNI/NEON) are selected at runtime in both, so within one binary you always get the fastest evaluation path your CPU supports.