brabo wrote on 06/04/19 at 09:05:37:

ordo elo (or self-play elo) can be very misleading, as I wrote a few posts before
can I ask what AMD card you're using? probably way better than mine

TopNotch wrote on 06/04/19 at 01:58:39:


OpenCL still gets reasonable performance on AMD cards. OpenCL on a CPU or an Intel iGPU is too slow. They just lack the horsepower.

It's mainly when comparing with the RTX cards that the gap widens up to ludicrous levels (i.e. an RTX 2060 is 15 times faster than a high end AMD card, whereas with a GTX 1060 it's only ~2.8).

I've tried this and essentially LC0 is utterly useless running on windows unless you have a Nvidia Graphics Card. I honestly don't understand why they even offer blas, opencl, its a joke.

Leon_Trotsky wrote on 06/02/19 at 22:37:00:

for Windows there are three versions (blas, opencl and cuda), did you try all three?
the blas and opencl versions are likely to be very slow, so don't expect too much

I followed the instructions, but still analysis window is totally blank 

update: I'm really not sure this weights file with 4305 elo is actually better than more recent ones about 3300
the first one gives some really weird evaluation in the following position, it initially thinks white is better there (the more recent file doesn't)



update2: Now I finally got it: self-play elo doesn't seem to be important, you have to look at https://docs.google.com/spreadsheets/d/1XSJiCcQpCLv0fNwrUn7jXjdkZFU63YFEWpdXv6dS... (I seem to remember this site was down when I first looked there)
So network number 42000 is strongest available for now

I am having a problem setting Leela on ChessBase 15. How did you make it work there to analyse  ¿

Leon_Trotsky wrote on 05/18/19 at 09:07:42:

did you follow the instructions at https://github.com/LeelaChessZero/lc0/wiki/Getting-Started ?
first you need a 'weights'-file from http://lczero.org/networks, presumably one with high elo, and then you need to open the lc0.exe file to initialise the configuration (this takes a while)
also important is to have the right drivers for the GPU (for Cuda or for Opencl), this took me a whole day to get this right at my new Ubuntu laptop, on an old Windows 7 laptop the lc0 program just crashed after a while
now that it's finally working I have to say it's pretty slow and I barely can move my mouse while it's analysing, I suppose you need a GPU with Cuda support (I have to use Opencl) for it to make any sense using Leela for analysis

edit: when I use 'blas' as backend instead of 'opencl' (so using CPU iso GPU), I can actually use my computer while analysing but the engine is of course even slower then
for info: I have AMD A6-9220 radeon r4 as processor with AMD Stoney as graphics card, so new but not very strong