LCZero

Client.exe is not needed to play. It is used if you want to support the project with the computing power of your computer.

AMD FX 8350 is a pretty modern processor, but it is much better to use the GPU for the LC0. I have a NVidia GT 1030 graphics card. It's a cheap card, but it's still 10x faster than a CPU (Intel i5 mobile or AMD Ryzen 3). For example, the difference is that on the CPU the depth of analysis is 3-4 and now 8-10.
Currently a good choice for the LC0 is the RTX card (much better than GTX). Thanks to it you can increase your speed by another 10x-20x.
If you think about OpenCL (AMD card), my observations show that the speed will be 50% of that of CUDA (NVidia). I checked both versions of LC0 on my GPU.

I downloaded:

lc0-v0.21.0-windows-cuda.zip

My path:

C:\Program Files (x86)\ChessBase\Engines\LC0_CUDA

in the directory:

cublas64_100.dll
cudart64_100.dll
cudnn64_7.dll
lc0.exe
weights_run1_42161.pb

I also use fritz and install lc0.exe as a normal chess engine. And that's it.

I don't change the engine parameters. The network file is detected automatically.

dear jaglo, my card graphic is very weak, Nvida gtx 650...

It's not bad. Your card has 67% of my card speed.

https://gpu.userbenchmark.com/Compare/Nvidia-GT-1030-vs-Nvidia-GTX-650/m283726vs3154

That's enough to do some research. It should still be faster than the CPU version.

Install lc0-v0.21.1-windows-cuda.zip

ok... Jaglo,.. I will try... I hope it works out...hug

Massimiliano Goi wrote:

Just made this mini-tournament... Lc0 was often in advantage just to drop all of it in the endings, most probably because it was short in time. Neural Network used: 44000.

deleted post

Last edited by jorge91 on Sat Jul 06, 2019 11:39 pm; edited 1 time in total (Reason for editing : deleted post)

Who is this?

why not introduce a uci leela version

Hello friends

I would like to understand where is located the experience file generated by the learning process

The best networks I think are:

32988

42848

J13B.2-136 (see results on CCC)

T60 training is in progress and I haven't checked it thoroughly yet.


They can be downloaded from here:

https://lczero.org/networks/

https://github.com/jhorthos/lczero-training/wiki/Leela-Training

Depends on hardware on single RTX i find this is very good if not the best
J20-460

this:
J13B.2-136 (see results on CCC)

On CCC it play on multiple GPUs 2080Ti and its need very good hardware.

Hello

Thanks for pointing the best networks to use on Lc0. There are so many...

What are the correct nodes to set for each of these networks?