Life In 19x19

roy7 wrote:It also looks like both gpus were using ~68% of gpu load.
I thought it should be something like 90% or more?

For max performance try using higher --threads and --batchsize than the defaults. Also once you've decided what threads/batchsize to use, run those settings with --full-tuner once.

At the beginning, when I run analyse mode without doing any moves, I see: visits per second are between 5000 and 30000.
Shouldn't it be somehow stable?

After 4000000 playouts, the visits are very low between 2000 and 5000.
Why?

When network evals hit the cache, they are still a visit in the ...

I was curious, so I was trying to run something similar yesterday. While I got a similar result, the games all looked identical up to move 40 or so.

I think I saw a comment in the github by someone else who ran comparisons that one of the tools people sometimes use to run these head to head ...

For those who might prefer it, there's an experimental "new" home page at http://zero.sjeng.org/home.

Pippen wrote:Why do we have it for Chess and not for Go?

Basically it's because there's way more interest in Chess than Go. Just compare the massive difference in size between the Discord channels. So that means a lot more hobby programmers available to make cool things like play.lczero.org.

LZ isn't ready for use for reviewing, it's still very weak.

The way Crazy Sensei handles the red/green thing is by having a graph making "made the same move Leela made" as 0, and then the bars are above or below zero by the amount better or worse than Leela. Makes it easy to instantly spot the big dips or better moves you made.

The author has released a public 0.11.0 beta here:

https://www.sjeng.org/leela-beta.html

And here's a change that could impact CGOS users:

The engine will now additionally ponder after kgs-genmove_cleanup commands (GTP version).

Also in Leela-0.10.5-16t1g - what 16t means? 16 threads of CPU? Which version of GPU it might be?
Just seems too big gap between others leela bots online.

16 threads. It's a Ryzen 1700 (no overclock). Running a 1080Ti GPU.

It runs on Manjaro/Arch version of linux. Linux has much lower system ...

aTan wrote:Almost all games lost because of wrong judgement of group statuses.

In Tromp Taylor rules every stone is alive. Dead stones of your opponent should be killed before you pass. This is why Leela needs the kgs-genmove_cleanup command which will behave that way.

What does --noponder do? Havne't find any clear explanation anywhere? Does it make leela stronger or weaker?

Leela will normally think during the opponent's time. You can turn that off using --noponder. Makes her weaker.

Leela does ponder by default, it might be the order the time_left and ...

aTan wrote:What does --noponder do? Havne't find any clear explanation anywhere? Does it make leela stronger or weaker?

Leela will normally think during the opponent's time. You can turn that off using --noponder. Makes her weaker.

Please note if you run Leela on cgos you need to approximate the Tromp Taylor rules by changing the cgos client to send kgs-genmove_cleanup instead of genmove. This way all dead stones will get removed.

For the 100 games I ran the only option was no pondering. I have a Ryzen 1700 and a 1080Ti. That ...

Leela is much stronger than the bad luck with a life/death issue makes it her look like sometimes.

Rayon also uses Nvidia gpus while Leela works best with AMD right?
For example, if you have amd you can't use rayon gpu mode.

Leela works fine either way. I believe at the moment the 1080Ti is the strongest GPU for Leela. Not sure on what Vega will do.

Rayon only works with Nvidia because of ...