Skip to content

Caissa v2.0

Latest
Latest

Choose a tag to compare

View all tags
@Witek902 Witek902 released this 19 Sep 17:37
· 5 commits to master since this release
2.0
bb72579

Caissa 2.0

This release switches the evaluation to a multilayer neural network and brings a new AVX-512 Ice Lake build, much better play at ultra-short time controls, and minor fixes.

Playing Strength

LTC, 60+0.6, UHO book:

ELO    40.41 +- 4.85 (95%)
CONF   60.0+0.60s Threads=1 Hash=128MB
GAMES  N: 5000 W: 1480 L: 901 D: 2619
PENTA  [5, 357, 1211, 908, 19]

https://furybench.com/test/7443/

STC, 10+0.1, UHO book:

ELO    29.15 +- 4.59 (95%)
CONF   10.0+0.10s Threads=1 Hash=16MB
GAMES  N: 6284 W: 1814 L: 1288 D: 3182
PENTA  [25, 550, 1478, 1052, 37]

https://furybench.com/test/7442/

STC, 10+0.1, DFRC UHO book:

ELO    28.69 +- 10.70 (95%)
CONF   10.0+0.10s Threads=1 Hash=16MB
GAMES  N: 1396 W: 398 L: 283 D: 715
PENTA  [13, 126, 314, 223, 22]

https://furybench.com/test/7445/

Changes

Evaluation

  • New multilayer network (32×768 → 1536) × 2 → 16 → 32 → 1:
    • The first layer is wider than before (1536 neurons per side, up from 1024).
    • Pairwise activation.
    • Two small hidden layers, with 8 output subnetworks selected by piece count.
    • Sparse inference skips the zero activations.
  • The network was trained from scratch for 312B positions, and its last layer was then tuned with SPSA.

Search & Time Management

  • MoveOverhead now actually reserves time. That's worth +28 to +63 Elo at time controls from 1+0 down to 0.1+0.001.
  • Tighter singular extension margin on exact TT bounds.
  • TT entries from earlier searches are always replaced.

Performance

  • New avx512icl build: +4.2% nps on Zen 4.
  • Fixed pext/pdep falling back to software emulation on Linux builds.
  • GCC codegen flags: +0.5% to +2.5% nps.

Bug Fixes

  • FEN parsing fixes (#25): X-FEN castling rights with several rooks, en passant validation, and rejection of malformed FENs.
  • hashfull reported 0 after 64 searches.

Tools

  • The CUDA trainer supports the multilayer net and gained an input factorizer, a cosine LR schedule and checkpoint/resume.
  • New utils permuteNet tool.
  • The legacy CPU trainer was removed.

Miscellaneous

  • The UCI_AnalyseMode option was removed.

Full Changelog: 1.26...2.0

Which binary version to choose?

Binary Typical CPUs
avx512icl Intel Ice Lake and newer, AMD Zen 4/5
avx512 Intel Skylake-X / Cascade Lake
bmi2 Intel Haswell and newer, AMD Zen 3
avx2 AMD Zen 1/2
sse4-popcnt / sse2 Older CPUs
Assets 8
Loading