Stefan Pohl Computer Chess

private website for chessengine-tests


Long thinking-time testing of LC Zero CUDA

 

Playing conditions:

 

Hardware: i7-6700HQ (2.6 GHz) Notebook (Quadcore), Windows 10 64bit, 12GB RAM

Fritzmark: 4 cores / 7 threads: 18.4 / 8825 (CPU-speed reduced to 99% (to switch off the Intel Turbo Boost))

GPU (used by LC Zero): Nvidia Cuda Gforce GTX 950M (4GB memory, grafics: 914 MHz, memory: 1.25 GHz). LC Zero CUDA calculates around 2600 rollouts/s in the starting position (measured with "go infinte"-command until depth 37), which is around 5x faster, than the normal LC Zero GPU version (!!)

Hash: 2 GB per engine

GUI: Fritz 16 GUI (no draw, no resign)

Tablebases: 3-5 Nalimov for GUI (GUI ends all games, when only 5 pieces remain on the chessboard)

Openings: 4 moves opening book. Download it here

Ponder, Large Memory Pages: Off

Thinking time: 12'+5'' per game/engine (average game-duration: 30'-40', so around 40-45 games per day are played). Average thinking time in the middlegame is around 20'' per move.

 

LC Zero Github (information, download Networks and LC Zero Engine): here

Interesting information about the LC Zero project on ChessBase.com: here

See a step-by-step manual for using LC Zero (and LC Zero CUDA) in FritzGUI and ShredderGUI here

 

Current participants: LC Zero 180619 (CUDA 9.2) Net432 plays a gauntlet vs. 10 opponents:

Stockfish 4 (Elo 3035), Andscacs 0.88 (Elo 3027), Houdini 1.5a (Elo 3023), Komodo 5 (Elo 3023), Critter 1.6a (Elo 3017), Nirvanachess 2.4 (Elo 3011), Hannibal 1.7 (2972), Texel 1.07 (Elo 2969), Laser 1.5 (Elo 2964), Protector 1.9.0 (Elo 2961).

Average opponent CEGT Elo (40moves/20'): 3000

All opponent-engines running with default settings and in singlethread-mode, like in CEGT-testruns.

 

But, to be sure about that, I will repeat the latest testrun (LC Zero 180602, Net 374) with LC Zero 180609, Net 374:

LC0 180602 played default with the settings above (FPU=0.2, Cpuct=3.1, Policy Softmax=1.0) and LC0 180609 plays default with Albert Silvers Clop-tuned settings (FPU=0.9, Cpuct=3.4, Policy Softmax=2.2), which should be stronger. I doubt that. But, when the testrun is finished, we will be sure (300 games with 12'+5'' vs. 10 opponents).

 

Because it turned out, that LC Zero 180609 was buggy, I aborted the repeated testrun. The result after 200 of 300 games was around -30 Elo weaker, but the reason could be the bug. But the new settings does not seem to be too bad, so from now, I will test LC Zero with the default-settings.

 

 

Latest update: 2018/06/15 - download all played games: here

 

 

LC0 Cuda(180602) N 374 (size: 15x192)(date: 180604): 300 (+ 86,=114,-100)
47.7 % (opponent-Elo: 3000) Elo-performance: 2984

 

LC0 Cuda(180530) N 359 (size: 15x192)(date: 180530): 300 (+ 79,=113,-108)
45.2 % (opponent-Elo: 3000) Elo-performance: 2967

 

LC0 Cuda(180521) N 322 (size: 15x192)(date: 180520): 300 (+ 68,=116,-116)
42.0 % (opponent-Elo: 3000) Elo-performance: 2944

 


lczero v0.8 (GPU) N 251 (size: 15x192)(date: 180505): 320 (+164,= 73,- 83)
62.7 % (opponent-Elo: 2524) Elo-performance: 2614


lczero v0.7 (GPU) N 214 (size: 10x128)(date: 180429): 300 (+157,= 85,- 58)
66.5 % (opponent-Elo: 2524) Elo-performance: 2643


lczero v0.7 (GPU) N 162 (size: 10x128)(date: 180421): 300 (+113,= 66,-121)
48.7 % (opponent-Elo: 2524) Elo-performance: 2515