Lc0 against SF14 one core

Discussion of computer chess matches and engine tournaments.

Moderators: hgm, Rebel, chrisw

Jouni
Posts: 3279
Joined: Wed Mar 08, 2006 8:15 pm

Lc0 against SF14 one core

Post by Jouni »

Code: Select all

                             
1   Stockfish 14        +63  +18/=82/-0 59.00%   59.0/100
2   Lc0 v0.27.0 69626   -63  +0/=82/-18 41.00%   41.0/100

1m + 1s and 8 moves book from SF framework.

Lc0 runs in GTX card with 2 knps and SF with 2 mnps. I am sure, that even with latest RTX card Lc0 has no chance vs SF multicore. You simply need A100 or V100 and still lose like in CCC and TCEC!
Jouni
User avatar
yurikvelo
Posts: 710
Joined: Sat Dec 06, 2014 1:53 pm

Re: Lc0 against SF14 one core

Post by yurikvelo »

SF gained 200+ Elo since NNUE introduction.
Leela kinda reached its ceiling.

Before NNUE, it was guesstimate who will win: HCE A/B or NN.
Now 200 elo frog leap leaves no chance to GPU computing.
Modern Times
Posts: 3546
Joined: Thu Jun 07, 2012 11:02 pm

Re: Lc0 against SF14 one core

Post by Modern Times »

yurikvelo wrote: Tue Sep 07, 2021 6:54 pm Now 200 elo frog leap leaves no chance to GPU computing.
For now. A new advance in GPU technology could tip the balance again ?
User avatar
yurikvelo
Posts: 710
Joined: Sat Dec 06, 2014 1:53 pm

Re: Lc0 against SF14 one core

Post by yurikvelo »

Modern Times wrote: Tue Sep 07, 2021 9:45 pm For now. A new advance in GPU technology could tip the balance again ?
Speed-up Leela development? Or speed-up nps for end-user of existing Leela networks?

Even if something extraordinary happen to nVidia and computational power of [4 x V100] will be available for consumer grade (something like single RTX 3090 with TDP=400W with power of 4 x V100) it will not cover already existing 200 elo gap.
Jouni
Posts: 3279
Joined: Wed Mar 08, 2006 8:15 pm

Re: Lc0 against SF14 one core

Post by Jouni »

In M.Y. big tournament SFdev 1 core was better than Lc0 RTX 2080TI. 1 cpu core beats 4352 cude cores + 2 cpu cores :!: . Is there any sense to continue Lc0 development?
Jouni
mehmet123
Posts: 670
Joined: Sun Jan 26, 2020 10:38 pm
Location: Turkey
Full name: Mehmet Karaman

Re: Lc0 against SF14 one core

Post by mehmet123 »

Jouni wrote: Fri Oct 01, 2021 6:15 pm In M.Y. big tournament SFdev 1 core was better than Lc0 RTX 2080TI. 1 cpu core beats 4352 cude cores + 2 cpu cores :!: . Is there any sense to continue Lc0 development?
It's only valid for fast games. Lc0 with 3060Tİ is stronger than Stockfish 13 with 8 cpu at long time control matches.


THE SSDF RATING LIST 2021-07-28
Rating + - Games Won Oppo
------ --- --- ----- --- ----
1 Lc0 0.26.3 Cuda(67362) 3060Ti 3581 41 -39 320 64% 3479
2 Stockfish 13 x64 1800X 3.6 GHz 3573 44 -40 320 70% 3428
3 Stockfish 12 NNUE x64 1800X 3.6 GHz 3568 32 -30 520 64% 3468
4 Dragon by Komodo x64 1800X 3.6 GHz 3557 34 -32 460 63% 3457
Jouni
Posts: 3279
Joined: Wed Mar 08, 2006 8:15 pm

Re: Lc0 against SF14 one core

Post by Jouni »

No. They used different books! So it was 40 games "book test" and ended 20-20. Serious testing only by using common book.
Jouni
User avatar
Rebel
Posts: 6991
Joined: Thu Aug 18, 2011 12:04 pm

Re: Lc0 against SF14 one core

Post by Rebel »

http://rebel13.nl/history.html#6

I got +53 for Lc0 v28
90% of coding is debugging, the other 10% is writing bugs.