Installing LC0 on CB GUI

[M ANSARI]

I finally decided to try out LC0 and managed to get some instructions to install it on the Chessbase GUI. I checked my VGA card and it is a Nvidia Quadro M4000 with around 1700 Cuda cores. So I installed the cuda version of LC0. I had some problems getting engine output but this got fixed when I renamed the weights file to a .txt extension. So it does seem it is working now but I am getting 14 kn/s in the start position. Not sure what I am supposed to get but that seems quite a bit low. Can someone give me an idea of what my output should be? My VGA card is pretty old but cost quite a lot when I got it as I need it for my 3D design renderings. Is there something wrong with my setup? Thanks in advance!

[Lion]

Not sure about your GPU but if of any help of comparison I get about 36 kn/s on a 2080ti which is the latest monster available....

Rgds

[AdminX]

Take a look at benchmarks:

https://github.com/LeelaChessZero/lc0/wiki/Benchmarks

https://docs.google.com/spreadsheets/d/ ... 1508569046

[MikeB]

I finally decided to try out LC0 and managed to get some instructions to install it on the Chessbase GUI. I checked my VGA card and it is a Nvidia Quadro M4000 with around 1700 Cuda cores. So I installed the cuda version of LC0. I had some problems getting engine output but this got fixed when I renamed the weights file to a .txt extension. So it does seem it is working now but I am getting 14 kn/s in the start position. Not sure what I am supposed to get but that seems quite a bit low. Can someone give me an idea of what my output should be? My VGA card is pretty old but cost quite a lot when I got it as I need it for my 3D design renderings. Is there something wrong with my setup? Thanks in advance!

Your Lc0 nps are very respectable, NN nodes are totally different. A very rough analogy is that for each node that is counted there was at a minimum at least game playout from that node ( and possibly many, many more) . So a NN node has much more information that a typical A/B engine search node. Anything over 10,000 nps is very, very good ( at least compared to what I have).

Here's good explanation of how it works using "tic-tac-toe" as the game instead of chess:
http://tim.hibal.org/blog/alpha-zero-ho ... -it-works/

It has been alleged that in the Stockfish 8 vs Alpha Zero match, that SF was seeing 70M nps per second and Alpha Zero was seeing around 80,000 NPS ( not start position , but average) - so just about a very rough 1000 to 1 ratio. You do have very a nice GPU.

[M ANSARI]

Thanks for all the information. I guess I have the program setup correctly and I really like the idea of using a VGA card to play chess. It is pretty remarkable how powerful and strong the engine plays !!! More impressive is how quickly the progress has been. I guess it is only a matter of time where the top engines will be using a similar format. Problem is now I really want to get a very powerful card.

[M ANSARI]

I did a quick test match between LC0 with latest downloadable software against SF 7. I expected an easy win for LC0 but SF 7 came out slightly ahead. I used no book to see how both engines would deal with things. I was really surprised that LC0 was a "pawn grabber" and lost quite a few games out of the opening due to grabbing poisoned pawns. SF 7 seemed to be much less materialistic. LC0 did win a few very nice games. I am guessing the the GPU hardware (14kns in start position) vs SF7 hardware (6 core CPU) maybe a factor? How is everyone using hardware equivalence between these 2 totally different platforms?

Now thinking more about it, maybe it is possible to build a hybrid engine that uses the MC search of the GPU to improve move choice of the AB engine? I remember a few years back this was an idea we had for Rybka but this never really came about.

[AdminX]

I did a quick test match between LC0 with latest downloadable software against SF 7. I expected an easy win for LC0 but SF 7 came out slightly ahead. I used no book to see how both engines would deal with things. I was really surprised that LC0 was a "pawn grabber" and lost quite a few games out of the opening due to grabbing poisoned pawns. SF 7 seemed to be much less materialistic. LC0 did win a few very nice games. I am guessing the the GPU hardware (14kns in start position) vs SF7 hardware (6 core CPU) maybe a factor? How is everyone using hardware equivalence between these 2 totally different platforms?

What Network ID were you using if I may ask? I have used 11258 and 32395. I prefer 11258 overall at the moment. I have heard that 32395 is stronger but its endgames while in winning positions leaves allot to be desired.

[M ANSARI]

I am using 35591. I figured that is the latest thus the best. I will try the 11258. Today I tried against SF9 and similar problem with taking poisoned pawns in the opening and also sometimes delays castling which causes problems.

[AdminX]

I am using 35591. I figured that is the latest thus the best. I will try the 11258. Today I tried against SF9 and similar problem with taking poisoned pawns in the opening and also sometimes delays castling which causes problems.

Yes, I understand and your assumption would be logical, however those tests are based on selfplay games and that is not often reliable. The latest is not always the best under those circumstances and Leela is known to have learning rate drops in selfplay games. A good source of info can be found on the LC0 (LCZero) Discord Channel (https://discordapp.com/). There they have a test-results sub channel that you can take a look at or ask questions in the general section. I normally use it to get a good idea of what Networks to work with.

[Guenther]

I am using 35591. I figured that is the latest thus the best. I will try the 11258. Today I tried against SF9 and similar problem with taking poisoned pawns in the opening and also sometimes delays castling which causes problems.

The 35 net is just a testbed for a future 40 net and has much less games than previous nets.
Best 10 nets e.g. 11258 or best 30 nets should be superior.

With 14kN/s vs. SF7 with 6 cores, LC0 should be better.

BTW watch the temperature of your GPU, it is necessary you have good enough cooling or it will be soon toast, with excessive testing