Lothar Jung schrieb:
Hier Test auf Discord:
**Match** Ceres v0.92 vs LC0 0.28rc1
**Network** j94-100
**LC0 options:** --nodes-as-playouts --smart-pruning-factor=0 --move-overhead=250 -w d:\weights\lczero.org\j94-100.pb.gz --backend=cuda-fp16 --backend-opts=gpu=3 --syzygy-paths="f:\sygyzy\5and6man"
**Ceres option:** None (defaults except the settings for the same GPUs, network, and tablebases as LC0)
**Hardware:** Intel 6142 Processor, each game on one GPU from pool of 2 A100, 2 Titan RTX
**Time control:** 50k nodes per move
**Software:** Integrated Ceres tournament manager on Windows 10
**Opening book:** TCEC18 + TCEC19
**Tablebases:** Syzygy 5 and 6-men
**Ordo:** ordo-win64 -Q -N 0 -D -a 0 -A "LC0_0" -W -n4 -V -s500 -J -p 50kCeresLC0.pgn
**Comment:** Fixed node test as suggested by Naphthalin to assess contribution of speed to Ceres performance.
**Comment:** 50k/move test here conveniently yields time and total nodes very close between Ceres and LC0 (within 3% each).
**Comment:** These results are with "out of the box" settings of both engines and are the ones I intended to run and post.
**Comment:** However subsequently I tried multiple variations of tweaking the batch sizing of one or both engines to improve results.
**Comment:** In these other tests, Ceres scored (+2, +20, +26, +29, +31) Elo.
```
# PLAYER : RATING ERROR POINTS PLAYED (%) CFS(%)
1 Ceres1 : 35 11 111.5 204 54.7 100
2 LC0_0 : 0 11 92.5 204 45.3 ---
White advantage = 90.39 +/- 11.56
Draw rate (equal opponents) = 86.39 % +/- 4.36
```
Hauptsache Ceres läuft auf einem M1 Chip und kann 7-Steiner Syzygy Tablebases verwenden.