LC Zero CUDA v0.16.0 Net 10555 (first test of a 20x256 Net from the testserver) finished.
Incredible Elo-jump: last testrun with a 15x192 Net had an Elo-perfomance of 3080. Net 10555 (20x256) has an Elo-performance of 3248 (+168 Elo). Means, that in the CEGT-ratinglist (singlecore, 40moves/20'), LC Zero would be now the world's #4 (Fire 7.1 has 3233 Elo in that list). Leela-ratio in my testruns is only 0.67 and the opponents are running in singlethread-mode, so the results are comparable to CEGT singlecore-ratings, I believe.
http://www.sp-cc.de
Here the single results of LC Zero CUDA v0.16.0 Net 10555:
vs.
Gull 3 x64 : 32 (+ 16,= 13,- 3), 70.3 %
Andscacs 0.94 : 32 (+ 9,= 19,- 4), 57.8 %
Ethereal 10.55 (PEXT) : 32 (+ 11,= 19,- 2), 64.1 %
Fizbo 2 : 32 (+ 17,= 13,- 2), 73.4 %
Deep Shredder 13 x64 : 32 (+ 8,= 19,- 5), 54.7 %
Booot 6.3.1_x64_popcnt : 32 (+ 11,= 20,- 1), 65.6 %
Fire 7.1 x64 popcnt : 32 (+ 8,= 18,- 6), 53.1 %
Komodo 11.3.1 64-bit : 32 (+ 7,= 14,- 11), 43.8 %
Houdini 6 x64-pext : 32 (+ 2,= 21,- 9), 39.1 %
Stockfish 9 64 BMI2 : 32 (+ 2,= 19,- 11), 35.9 %
(Perhaps you have to clear your browsercache or reload the website)
SPCC: Testrun of LC Zero Net 10555 finished
Moderators: hgm, Rebel, chrisw
-
- Posts: 2435
- Joined: Sat Sep 03, 2011 7:25 am
- Location: Berlin, Germany
- Full name: Stefan Pohl
-
- Posts: 358
- Joined: Thu Jan 22, 2015 3:21 pm
- Location: Zurich, Switzerland
- Full name: Jonathan Rosenthal
Re: SPCC: Testrun of LC Zero Net 10555 finished
So a quick question with respect to the Leela ratio. I assume for the Leela ratio you calculated your GPU vs a single core?
I have no doubt you reported your results correctly, to me it is just a bit confusing since so far my impression was that people were defining their Leela Ratio as fixed for a CPU/GPU setup, but technically this is only true if you give your program all your cores. I think sometimes it may make sense to compare different engines with differing numbers of cores with NN engines, in which case technically different Leela Ratio numbers are needed. To fix this ambiguity, maybe we should introduce a Leela Per Core Ratio? Does SF NPS scale completely linearly per core?
In any case, Leela progress is interesting to follow, nice testing =)
I have no doubt you reported your results correctly, to me it is just a bit confusing since so far my impression was that people were defining their Leela Ratio as fixed for a CPU/GPU setup, but technically this is only true if you give your program all your cores. I think sometimes it may make sense to compare different engines with differing numbers of cores with NN engines, in which case technically different Leela Ratio numbers are needed. To fix this ambiguity, maybe we should introduce a Leela Per Core Ratio? Does SF NPS scale completely linearly per core?
In any case, Leela progress is interesting to follow, nice testing =)
-Jonathan
-
- Posts: 2435
- Joined: Sat Sep 03, 2011 7:25 am
- Location: Berlin, Germany
- Full name: Stefan Pohl
-
- Posts: 1346
- Joined: Sat Apr 19, 2014 1:47 pm
Re: SPCC: Testrun of LC Zero Net 10555 finished
So this net or another one could be a good opponent as well for your Stockfish testing.