Discussion of anything and everything relating to chess playing software and machines.
Moderators: bob, hgm, Harvey Williamson
Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
-
Glarean
- Posts: 191
- Joined: Sun Oct 05, 2008 11:04 am
- Location: Switzerland
- Full name: Walter Eigenmann
-
Contact:
Post
by Glarean » Fri May 22, 2020 6:40 pm
In a tournament with 17 Engines, I investigated whether Leela & Co. or AlphaBeta played better.
The trend is toward NN, but at the moment we're still tied.
Hardware/Software: AMD Ryzen7 2700X 3,7 GHz - 1024 Mb Hash - 8Cores/16Threads - GPU RTX2080 - GUI Fritz 17 - 4min & 2sec TC - 5-Moves-Opening-Book - 5-men-Syzygy-Tablebases
The Leela victories are very impressive. There are numerous positional highlights.
Above all Leela can handle the knights with virtuosity.
One of four examples:
3qkb1r/1r3pp1/1nn1p2p/p2pP2P/1ppP4/1PP2NR1/P2BNPP1/1R1Q2K1 w k
Ne1-c2-e3 !
The whole report and the download files (CBH/PGN/EPD) here:
https://glarean-magazin.ch/2020/05/22/c ... urnaments/
Greetings: Walter
-
Glarean
- Posts: 191
- Joined: Sun Oct 05, 2008 11:04 am
- Location: Switzerland
- Full name: Walter Eigenmann
-
Contact:
Post
by Glarean » Fri May 22, 2020 7:45 pm
Thanks for your feedback.
Yes, it's the age-old struggle at Prokofiev and Anand and everywhere:
Knight against Knight...
Greetings: Walter
-
brianr
- Posts: 397
- Joined: Thu Mar 09, 2006 2:01 pm
Post
by brianr » Fri May 22, 2020 8:02 pm
Lc0 and most of the other NN engines should be run with 2 threads, not 1 (or number-of-GPUs + 1).
Incidentally, more than that actually hurts performance (very unlike A/B engines).
-
Dann Corbit
- Posts: 10836
- Joined: Wed Mar 08, 2006 7:57 pm
- Location: Redmond, WA USA
-
Contact:
Post
by Dann Corbit » Fri May 22, 2020 8:39 pm
The PGN archive also contains ChessBase format.
Taking ideas is not a vice, it is a virtue. We have another word for this. It is called learning.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.
-
Glarean
- Posts: 191
- Joined: Sun Oct 05, 2008 11:04 am
- Location: Switzerland
- Full name: Walter Eigenmann
-
Contact:
Post
by Glarean » Fri May 22, 2020 9:55 pm
You're right, thanks for the advice!
Corrected.
-
Glarean
- Posts: 191
- Joined: Sun Oct 05, 2008 11:04 am
- Location: Switzerland
- Full name: Walter Eigenmann
-
Contact:
Post
by Glarean » Fri May 22, 2020 10:01 pm
brianr wrote: ↑Fri May 22, 2020 8:02 pm
Lc0 and most of the other NN engines should be run with 2 threads, not 1 (or number-of-GPUs + 1).
As far as I know, it's not important, one core is enough. The calculation of the NN engines is done almost exclusively via the GPU. BTW, a very strong Nvidia-RTX2080 was used here...
-
brianr
- Posts: 397
- Joined: Thu Mar 09, 2006 2:01 pm
Post
by brianr » Fri May 22, 2020 10:38 pm
Glarean wrote: ↑Fri May 22, 2020 10:01 pm
brianr wrote: ↑Fri May 22, 2020 8:02 pm
Lc0 and most of the other NN engines should be run with 2 threads, not 1 (or number-of-GPUs + 1).
As far as I know, it's not important, one core is enough. The calculation of the NN engines is done almost exclusively via the GPU. BTW, a very strong Nvidia-RTX2080 was used here...
Exactly 2 threads should be used for Lc0 with one GPU (which is why it is the default value)
This turns out to be the optimal way to collect batches of work to keep the GPU busy.
More than two incurs locking penalties (currently).
Less than two is unlikely to feed enough work to the GPU.
The speed of the CPUs is somewhat important, with higher clock rates being better.
Up to you, of course.
Last edited by
brianr on Fri May 22, 2020 10:43 pm, edited 1 time in total.
-
Glarean
- Posts: 191
- Joined: Sun Oct 05, 2008 11:04 am
- Location: Switzerland
- Full name: Walter Eigenmann
-
Contact:
Post
by Glarean » Fri May 22, 2020 10:42 pm
I'd be interested: Any programming or statistical evidence to support that claim?
