I noticed that some pc come with options of memory size for the gaming card. I'm not talking about ram.
For example, GTX 1650 (4GB vs 8GB)
Would more dedicated memory to the chip make a significant difference (or any) to Leela's playing strength?
Is more dedicated memory to gaming chip better for Leela
Moderators: hgm, Rebel, chrisw
-
- Posts: 508
- Joined: Fri Jun 04, 2010 7:23 am
-
- Posts: 536
- Joined: Thu Mar 09, 2006 3:01 pm
Re: Is more dedicated memory to gaming chip better for Leela
Short answer for game playing is no.
Longer answer is that larger nets which need more GPU RAM will take longer on your GPU whereas smaller nets will see more nodes. You can experiment, but a lot will depend on time controls. Even very large 30x384 nets fit it 2GB but will be quite slow. See:
https://github.com/LeelaChessZero/lc0/w ... ts-for-Lc0
For actually training a net, more GPU RAM will be needed, although that can be mitigated with more batch splits. The other category is contributing by using the client to play self-play "training games". These are nodes-based time controls and should be OK but again larger nets will be too slow on that card to be worthwhile.
Longer answer is that larger nets which need more GPU RAM will take longer on your GPU whereas smaller nets will see more nodes. You can experiment, but a lot will depend on time controls. Even very large 30x384 nets fit it 2GB but will be quite slow. See:
https://github.com/LeelaChessZero/lc0/w ... ts-for-Lc0
For actually training a net, more GPU RAM will be needed, although that can be mitigated with more batch splits. The other category is contributing by using the client to play self-play "training games". These are nodes-based time controls and should be OK but again larger nets will be too slow on that card to be worthwhile.
-
- Posts: 708
- Joined: Mon Jan 16, 2012 6:34 am
Re: Is more dedicated memory to gaming chip better for Leela
Very minimal (5 to 20 elo) in bullet and blitz. A bit more advantage for long analysis.
Definately a lot of increasing strength ( approx 100 elo, depends on net and time control), if you consider 2060 RTX instead. (approx 4x faster, depends on nets )