Posts : 2608 Join date : 2020-11-17 Location : Netherlands
Subject: Lc0 question Fri Nov 27, 2020 12:38 pm
In ProDeo 3.0 I have analyzed the main book with SF12 but as noted elsewhere the SF12 analysis procedure switches from normal to NNUE after each position causing that the scores are not in sync. And so I am in the process using Lc0 instead.
Currently (as an investigation) I have tried 2 nets.
1. An 128x10 net (size 6.21 Mb) depth=7 maxtime 1000ms.
2. An 384x30 net (size 130 Mb) depth=5 maxtime 1000ms.
Both use about the same time to finish.
The question is, which net performs best when it is about (say) the first 10 moves of a game.
mwyoung
Posts : 880 Join date : 2020-11-25 Location : USA
Subject: Re: Lc0 question Fri Nov 27, 2020 4:11 pm
Admin wrote:
In ProDeo 3.0 I have analyzed the main book with SF12 but as noted elsewhere the SF12 analysis procedure switches from normal to NNUE after each position causing that the scores are not in sync. And so I am in the process using Lc0 instead.
Currently (as an investigation) I have tried 2 nets.
1. An 128x10 net (size 6.21 Mb) depth=7 maxtime 1000ms.
2. An 384x30 net (size 130 Mb) depth=5 maxtime 1000ms.
Both use about the same time to finish.
The question is, which net performs best when it is about (say) the first 10 moves of a game.
Hello Ed,
I have many thousands of games testing Lc0. The bigger nets with all else being equal perform the best. The best Lc0 nets are found here. https://github.com/jhorthos/lczero-training/wiki/Leela-Training The J94-80 net is new and most likely the best. And I have tested the other nets. And they are also best.
The J92 nets are also about equal. And they are Zero trained.
Admin likes this post
kai_laskos
Posts : 37 Join date : 2020-11-17
Subject: Re: Lc0 question Fri Nov 27, 2020 7:13 pm
Admin wrote:
The question is, which net performs best when it is about (say) the first 10 moves of a game.
First ten moves are mostly positional and at about 1s/move with an RTX GPU clearly 384x30 nets win. I would choose the last zero net J92-330.
Admin likes this post
Admin Admin
Posts : 2608 Join date : 2020-11-17 Location : Netherlands
Subject: Re: Lc0 question Mon Nov 30, 2020 9:59 am
Analyzed the first 10 moves of the ProDeo book with Lc0, the rest remains SF12 and I am getting much better results. Amazing.
Chris Whittington
Posts : 1254 Join date : 2020-11-17 Location : France
Subject: Re: Lc0 question Mon Nov 30, 2020 10:31 am
Admin wrote:
Analyzed the first 10 moves of the ProDeo book with Lc0, the rest remains SF12 and I am getting much better results. Amazing.
Not sure if you tried it, but when I analysed positions en masse with lc0, it was possible to launch several processes without affecting nps too badly. Which is weird because each process has to share the same GPU. My 2080 is on a PC with only six cores, so no idea if your 20 core would scale
Admin Admin
Posts : 2608 Join date : 2020-11-17 Location : Netherlands
Subject: Re: Lc0 question Mon Nov 30, 2020 10:52 am
I tried that too with my 1060, system froze.
I can do 20 (or so) threads with the CPU edition, but the 1060 is faster.
Chris Whittington
Posts : 1254 Join date : 2020-11-17 Location : France
Subject: Re: Lc0 question Mon Nov 30, 2020 12:09 pm
Admin wrote:
I tried that too with my 1060, system froze.
I can do 20 (or so) threads with the CPU edition, but the 1060 is faster.
Will it work for two processes? Should give same nps but twice as many games. For N processes there will be a bottleneck caused by wanting too much GPU RAM, 1060 obvs not optimal. I think you can tell LC0 to limit down GPU RAM usage.
Admin Admin
Posts : 2608 Join date : 2020-11-17 Location : Netherlands
Subject: Re: Lc0 question Mon Nov 30, 2020 1:33 pm
Won't matter, unlike your 2080 my 1060 uses 100% GPU time.
kai_laskos
Posts : 37 Join date : 2020-11-17
Subject: Re: Lc0 question Mon Nov 30, 2020 2:01 pm
Chris Whittington wrote:
Admin wrote:
Analyzed the first 10 moves of the ProDeo book with Lc0, the rest remains SF12 and I am getting much better results. Amazing.
Not sure if you tried it, but when I analysed positions en masse with lc0, it was possible to launch several processes without affecting nps too badly. Which is weird because each process has to share the same GPU. My 2080 is on a PC with only six cores, so no idea if your 20 core would scale
Hmm, that's strange. My GPU utilization with single Lc0 instance on RTX 2070 is 95-99% if the time used per move or analysis is higher than several dozen milliseconds, that is, most of the time. So, I never use multiple instances of Lc0.
mwyoung
Posts : 880 Join date : 2020-11-25 Location : USA
Subject: Re: Lc0 question Mon Nov 30, 2020 2:17 pm
Chris Whittington wrote:
Admin wrote:
Analyzed the first 10 moves of the ProDeo book with Lc0, the rest remains SF12 and I am getting much better results. Amazing.
Not sure if you tried it, but when I analysed positions en masse with lc0, it was possible to launch several processes without affecting nps too badly. Which is weird because each process has to share the same GPU. My 2080 is on a PC with only six cores, so no idea if your 20 core would scale
Yes. I do know why this is true. But you also see this behavior when test matching 2 Lc0 engines and nets. And turn Ponder on for the match. It is crazy.
Home 
