Skip to content

Triumviratus - 7.0

Latest
Latest

Choose a tag to compare

View all tags
@Tors3 Tors3 released this 10 Sep 19:28
· 3 commits to main since this release
v7.0
933f225

UCI chess engine · NNUE · GPLv3 — by Francesco Torsello, in collaboration with Maurizio Platino.
Development log: DEVELOPMENT_7.0.md ·
networks: NETWORKS.md

Strength

Opponent TC Games Elo
Triumviratus 6.0 25+0.25 3,000 +21.34 ± 6.66
Triumviratus 6.0 5+0.05 9,000 +25.18 ± 4.28
Stormphrax 8.0.0 25+0.25 1,000 +50.03 ± 11.93
Hobbes 3.0 25+0.25 1,000 +47.19 ± 12.19
Cinder 0.6.1 25+0.25 1,000 −9.73 ± 11.42

1 thread, UHO 2024 (+0.85/+0.94), the same instruction set on both sides. An unbalanced book
widens the gaps: these figures compare engines under these conditions, they do not predict a rating.

What's new

  • legio-septima, the first network trained from scratch with base and feature blocks together:
    TRANN2, Stockfish's SFNNv16 feature set plus our own PassedPawns block.
  • Evaluation constants retuned for this network: +3.75 ± 3.49 Elo.
  • +5% NPS at identical node counts: release builds compile the tuning parameters as constants.
  • score cp recalibrated for this network: +100 means a 50% chance of winning.
  • Hash up to 65536 MB, Threads up to 512.

Which build to download

Your CPU Build
AMD Zen 5 — Ryzen 9000, EPYC 9005 avx512icl
AMD Zen 4 — Ryzen 7000 desktop, Ryzen 8000G, EPYC 9004 vnni512
AMD Zen 3 — Ryzen 5000 desktop, EPYC 7003 avx2
AMD Zen 1 / Zen 2 — Ryzen 1000 to 3000, EPYC 7001 / 7002 avx2-nopext
Intel Sapphire Rapids and later — Xeon Scalable 4th gen and later avx512icl
Intel Ice Lake, Tiger Lake, Rocket Lake — Core 11th gen, Core 10th gen G models avx512icl
Intel Cascade Lake, Ice Lake SP — Xeon Scalable 2nd / 3rd gen, Core X 10th gen vnni512
Intel Skylake-X / SP — Xeon Scalable 1st gen, Xeon W-21xx, Core X 7th to 9th gen avx512
Intel Haswell to Comet Lake, Alder Lake and later — Core 4th to 10th gen, 12th to 14th gen, Core Ultra avx2-intel

Windows: Triumviratus_7.0_<build>.exe. Linux: triumviratus_linux_<build>, with avx512,
avx2, avx2-intel and avx2-nopext; on any AVX-512 CPU use avx512. Linux builds need glibc
2.38 or newer. Ryzen laptop model numbers mix generations: go by the architecture. Android arm64:
triumviratus_android_dotprod on cores with the dot-product extension (ARMv8.2 and later),
triumviratus_android otherwise.

The network is embedded in every binary; nn-legio-septima.nnue is only needed to override it.
bench prints 240503 on every binary. Checksums in SHA256SUMS.txt.

GPLv3. The NNUE inference code derives from Stockfish; the PawnPair feature design is Jonathan
Hallström's (Pawnocchio); Syzygy probing uses Fathom. Full credits in the README.

Assets 16
Loading