Hier ein Test auf Discord mit starker Hardware:
**Gauntlet:** SF-NNUE vs. LC0 (tuned parameters)
**LC0 version:** v0.26.3
**LC0 options:** Network: J92-330, Backend=cuda-fp16, Threads=1, MLH: TCEC-SuFi S18, MinibatchSize=64, MaxPrefetch=32, MaxCollisionEvents=32, MaxOutOfOrderEvalsFactor=2.0, SmartPruningMinimumBatches=2, MoveOverheadMs=0, ScoreType=centipawn_2019, RootHasOwnCpuctParams=false
**SF options:** SF-dev, Threads=2, Hash=8, Contempt=0, "Move Overhead"=0, SyzygyProbeDepth=10
**Hardware:** Ryzen 9 3950X (3.5GHz) + RTX 3080
**Time control:** 1.8s/game+0.03s/move (LC0), 3.0s/game+0.05s/move (SF), time forfeits disabled
**Speed:** ≈1000 nodes/move (LC0)
**Book:** unbalanced 3-move book
**Tablebases:** 6-man
**Adjudication:** 6-man TBs + -resign movecount=3 score=500, -draw movenumber=20 movecount=5 score=10
**Software:** Cutechess-CLI
**LC0 options:** tuned: CPuct=1.350, FpuValue=0.228, PolicyTemperature=1.497, CPuctFactor=3.468, CPuctBase=20253
```
tuned default
CPuct 1.350 2.147
FpuValue 0.228 0.443
PolicyTemperature 1.497 1.607
CPuctFactor 3.468 2.147
CPuctBase 20253 18368```
``` # ENGINE : RATING ERROR CFS(%) GAMES DRAWS(%) OppN
1 lc0.net.J92-330-tuned : 10.4 6.9 99.8 4500 55.4 1
2 stockfish-dev : 0.0 ---- 63.3 9000 55.8 2
3 lc0.net.J92-330 : -1.2 6.8 --- 4500 56.2 1
White advantage = 47.1 +/- 2.4
Draw rate (equal opponents) = 57.0 % +/- 0.5```With default lc0 as anchor:
``` # ENGINE : RATING ERROR CFS(%) GAMES DRAWS(%) OppN
1 lc0.net.J92-330-tuned : 11.6 9.8 99.8 4500 55.4 1
2 stockfish-dev : 1.2 6.8 63.3 9000 55.8 2
3 lc0.net.J92-330 : 0.0 ---- --- 4500 56.2 1
White advantage = 47.1 +/- 2.4
Draw rate (equal opponents) = 57.0 % +/- 0.5```**Comment:** CPuct/FpuValue/PolicyTemperature from tuning result <
https://discord.com/channels/425419482568196106/766262364949315614/775389096930836500>. CPuctFactor/CPuctBase from T70 tuning <
https://discord.com/channels/425419482568196106/766262364949315614/775073169479696394>, but not relevant for this TC.
Lothar