Page 1 of 2
Komodo?
Posted: Mon Oct 08, 2018 1:39 pm
by Hurnavich
Hi, anyone know what is the development situation with Komodo? i think the last update i got as a subscriber at the time was 130518
Been a long time no news.Just asking anyone in the know.
Re: Komodo?
Posted: Mon Oct 08, 2018 1:49 pm
by Werewolf
+1 Also wondering
Re: Komodo?
Posted: Mon Oct 08, 2018 6:23 pm
by lkaufman
Should be soon. Main delay has been getting MCTS version to be able to use more than 8 threads profitably, but I think we're close to solving that problem.
Re: Komodo?
Posted: Mon Oct 08, 2018 6:30 pm
by S.Taylor
And THEN it will wipe the floor with sf!?
Re: Komodo?
Posted: Mon Oct 08, 2018 7:34 pm
by lkaufman
Only small gains for normal Komodo, but huge gains for MCTS version. The normal version will still be stronger, but the MCTS version will probably be stronger when using MultiPV of 3 or more. Barring some major discovery, the only chance to catch Stockfish in the next year is with the MCTS version.
Re: Komodo?
Posted: Mon Oct 08, 2018 7:39 pm
by S.Taylor
So MCTS is indeed overtaking the normal?
Re: Komodo?
Posted: Mon Oct 08, 2018 8:17 pm
by lkaufman
Well it has a way to go to catch up with normal Komodo (except when both are using MPV 3 or more, which hurts normal Komodo badly but doesn't hurt Komodo MCTS at all), but we've probably cut the gap in half in three months, so you can project for yourself what this would mean if it continues.
Re: Komodo?
Posted: Mon Oct 08, 2018 8:19 pm
by S.Taylor
That would be great!
Re: Komodo?
Posted: Mon Oct 08, 2018 10:39 pm
by Werewolf
lkaufman wrote: ↑Mon Oct 08, 2018 8:17 pm
we've probably cut the gap in half in three months, you can project for yourself what this would mean if it continues.
It would mean infinitely catching up the normal version
Do you have any more data on MCTS scaling compared with alpha beta? Are they still looking pretty similar on high core counts (say 8)?
Re: Komodo?
Posted: Tue Oct 09, 2018 12:24 am
by lkaufman
MCTS scaling is close to alpha-beta scaling (talking about Komodo dev.) for up to four threads, somewhat worse from 4 to 8, but no benefit at all beyond 8. This number was 15 at last release, but it seems that the huge improvements since then on 1 to 4 threads weren't fully matched at higher counts. I want to raise this number back up to something more reasonable than 8 before release, otherwise I'd say we are ready now. Does anyone know whether Leela scales ok up to high thread counts (compared with say stockfish)? I don't follow the Leela news that closely. I suppose it scales well with no GPU but can't really use many CPUs with a good GPU. Is that about right?