Lc0 NN329 against Houdini 4 in 2 LTC games

[Guenther]

Total played 2 games, both won by Lc0, side and reversed from standard opening position. It took 5-6 hours of this evening and night to complete them. The games are here:

https://lichess.org/IBFHR8op
https://lichess.org/9Rm9ZUeP

BTW, if you dislike the behaviour of LC0 to just report min-depth as depth in the output, which is saved to pgn,
then you can correct it by using inbetween. (server2client part, the others according to your installation/taste)

It seems I can run the newer lc0 versions too now (cuda 9.2), only the first game always stalls in cc-cli, after that it runs smoothly.

e.g.

Code: Select all

[InBetween]
Debug := 3
Log := LC0_329.log
CommandLine := lc0.exe -w ID329
 
[Client2Server]
; Set transalation of commands from the client here
; format word := translation
 
[Server2Client]
; Set transalation of commands from the server here
; format word := translation
info depth 1 seldepth := info depth
info depth 2 seldepth := info depth
info depth 3 seldepth := info depth

[stavros]

according to kai tests and with a more fast gpu HW (4x1080i or nvidia v100 8000$ )lc0 already caught sf8 witth a cheaper hardware than A0
with 4 TPUsv2 maybe is better than sf9 too.seems lczero training NN is better than A0.

[Werewolf]

Thanks you for the games mr Laskos. You might add some core for Houdini, 1 core is little, even if Houdini 4 was the top engine for a long time.

So Cudnn is now working without bug ?

The CUDA engine is a a different engine than the master v.10, but by now it seems significantly stronger (by at least 100 Elo points in my estimate). Its speed is factors of 6-8 higher in NPS.

I've installed CUDA 9.2, cuDNN 7.14, etc. and finally got it working.

Is it normal to see massive fluctuations in speed? I'm running under Arena. In the opening it's 4-8K but in the middlegame drops to 1-2K. 1060 6GB card being used.

[shrapnel]

Is it normal to see massive fluctuations in speed?

Yes, it is.
There is also some relation to the lc0 Parameters being used (default or not) and the particular binary you are using.

[shrapnel]

In the opening it's 4-8K but in the middlegame drops to 1-2K. 1060 6GB card being used.

Finally got down to a little overclocking.
I'm hitting 23k Speeds !

Stop wasting your time with the tiddlywinks 1060, Werewolf bro, go for the 1080 Ti.
I know you can !
lc0 is holding the latest SDVs to draws, on my System in 1min/move games.
Just a matter of time before it leaves Stockfish/Komodo/Houdini in the Dust !
But for that, a STRONG GPU is needed.

[Lion]

Impressive!

Can you post games of such games?

What is the CPU on which SF9 is playing?

rgds

[shrapnel]

Impressive!

Can you post games of such games?

What is the CPU on which SF9 is playing?

rgds

Thanks !
I will post a few games later on.
What CPU ? See my Signature.

[Lion]

Thanks!

Looking forward to see the games!

rgds

[Jhoravi]

Thanks !
I will post a few games later on.
What CPU ? See my Signature.

That's great! It's interesting to see a similar setup to Laskos wherein there are two games at long time control.

[yanquis1972]

Thanks you for the games mr Laskos. You might add some core for Houdini, 1 core is little, even if Houdini 4 was the top engine for a long time.

So Cudnn is now working without bug ?

The CUDA engine is a a different engine than the master v.10, but by now it seems significantly stronger (by at least 100 Elo points in my estimate). Its speed is factors of 6-8 higher in NPS.

I've installed CUDA 9.2, cuDNN 7.14, etc. and finally got it working.

Is it normal to see massive fluctuations in speed? I'm running under Arena. In the opening it's 4-8K but in the middlegame drops to 1-2K. 1060 6GB card being used.

yeah, it's quite different form a/b engines. i think the reported averages tend to be overstated, but cranking up the CPUCT parameter gives a large increase in NPS (near double, it think??). probably some other variables too.

you can view benchmarks here https://docs.google.com/spreadsheets/d/ ... =857482380 & your #s are right in-line.