Apparently Ceres results are sensitive to the CPU/GPU ratio. It does well (relatively to Lc0) on a weak CPU, and strong GPU. My hardware can be seen as being somewhat "Ceres friendly". On the other hand, Ceres is supposed to perform better relatively to Lc0 with smaller nets (such as 591226, 703810 etc.), which wasn't the case here (66740 is a large 30x384 net).
That being said, results were very close between the two engines (all games vs Stockfish).
Match: Lc0 26.3 and CERES vs Stockfish 201225 bmi2 (18T)
Hardware: RTX 3080, i9-10900kf
Time control: 120s + 2s
Openings: TCEC superfinals 10-19 (500 pos.) with a few modifications.
Ceres sysbench: CPU = 121, nps/batch = 15 392
bench (startpos. 15s): Stockfish = 16,500,000 nps / Lc0 = 24 300 nps / CERES = 22 725 nps
Stockfish: threads=18, hash=2048
Lc0: cuda-fp16, nncache=10,000,000, threads=2, minibatch-size=204, mlh=on
mlh tcec-19: moves-left-max-effect=0.2, moves-left-threshold=0, moves-left-slope=0.004, moves-left-scaled-factor=1, moves-left-quadratic-factor=0, moves-left-constant-factor=0
TBs and adj.: syzygy 5-men, resign movecount=5 score=600, draw movenumber=30 movecount=5 score=5
Games: https://gofile.io/d/kA1hWM
Code: Select all
# PLAYER : RATING ERROR PLAYED (%) CFS W D L D(%)
1 stockfish-201225-18T : 0.0 ---- 2000 52.08 99 288 1507 205 75.35
2 lc0-263-66740 : -12.0 10.2 1000 48.35 80 107 753 140 75.30
3 CERES-66740 : -18.1 10.3 1000 47.50 --- 98 754 148 75.40
White advantage = 64.09 +/- 3.79
Draw rate (equal opponents) = 82.07 % +/- 1.08
Code: Select all
Engine Depth MIDG EARLY ENDG LATE
CERES-66740 6.53 8.22 | 7.71 | 6.59 | 6.05
lc0-263-66740 12.48 14.39 | 14.48 | 11.13 | 8.17
stockfish-201225-18T 40.50 34.46 | 35.23 | 43.06 | 56.54