Consider changing your CPUCT value for different time controls. 3.4 is good for bullet chess. 2.0 is good for TCEC controls (120’/40).
If you have “only” one video card, set threads to 2. More doesn’t help.
Lc0 settings for Leela on GPU- Nvidia RTX 2080
Moderators: hgm, Rebel, chrisw
-
sincerelybrent
- Posts: 47
- Joined: Fri Feb 15, 2019 11:41 pm
- Full name: Brent M
-
ankan
- Posts: 77
- Joined: Sun Apr 21, 2013 3:29 pm
- Full name: Ankan Banerjee
Re: Lc0 settings for Leela on GPU- Nvidia RTX 2080
I agree that 2 threads is optimal for single GPU, and you must have --backend=cudnn-fp16 (to get ~2.5X speedup over the default cudnn fp32 backend).
Changing some of the settings for more NPS can weaken the search. E.g, I wouldn't suggest increasing minibatch-size more than the default (256).
Increasing NNCache setting is pretty safe though (i.e, increases NPS without changing the search). I normally set it to 10million and leave all other settings to defaults. People have found some small improvements with non-default CPUCT settings, but I haven't personally tested them.
Changing some of the settings for more NPS can weaken the search. E.g, I wouldn't suggest increasing minibatch-size more than the default (256).
Increasing NNCache setting is pretty safe though (i.e, increases NPS without changing the search). I normally set it to 10million and leave all other settings to defaults. People have found some small improvements with non-default CPUCT settings, but I haven't personally tested them.