LCZero 0.20.2 Cuda (32930) @ CEGT 5'+3" ponder=on

[ThatsIt]

Hi !

The test is finished:

http://www.cegt.net/5Plus3Rating/Pureli ... liste.html

Code: Select all

1. Stockfish 10.0 x64 1CPU     3440 [+- 16] out of 1300 games
2. LCZero 0.20.2 Cuda (32930)  3415 [+- 15] out of 1300 games
3. Houdini 6.0 x64 1CPU        3352 [+- 14] out of 1300 games
4. Komodo 12.2.2 x64 1CPU      3349 [+- 13] out of 1300 games
...

Best wishes,
G.S.
(CEGT team)

[Hugo]

Thanks for the test !

1300 games. WOW.

I did 5min + 3 sec ponder on with my RTX 2060 downclocked to 1400MHz.(NN 41812)
Opponent was SF 5th April2019 with 13 cpu (E5-2697A v4).
Leela ratio 1,06
100 games, TCEC 10 final startpossitions
Stockfish won, match ended +16 / -12 / =72
C.K.

[Jouni]

SF 5.4.2019 was buggy and -100 ELO to SF10!

[Hugo]

Thx for pointing this out. I may switch to 9th April version.

were did you get those informations please?
Do you have something concrete?

could not find any information about that.

kind regards

C.K.

SF 5.4.2019 was buggy and -100 ELO to SF10!

[Jouni]

Here https://nextchessmove.com/dev-builds.

[Jouni]

Bug was interesting. It distroyed time management badly, but in test positions you can't see anything wrong! It was MC formatting error fixed 6.4.

[Hugo]

Bug was interesting. It distroyed time management badly, but in test positions you can't see anything wrong! It was MC formatting error fixed 6.4.

I did some pre-testing before I started the 5m+3s matches. I noticed the bad time management, but was not really wondering, bcs it happens often in Fritz Gui that SF needs some slow mover adjustment.
So this time, SF had a good time mangement with slow mover set to 20.
btw. for Lc0 I have set slow mover to 20000, time usage on both engine is very similar now.

thx for the link to intresting website ,
regards,
C.K.