Lc0 NN329 against Houdini 4 in 2 LTC games

[Laskos]

From the standard opening position of Chess, I played two relatively long time control games, side and reversed, of LC0 and Houdini 4, a former top dog, and a 3241 CCRL 40/4' Elo rated engine on one thread.
Conditions are the following.
I used CUDA 9.2 version of GPU Lc0 with NN329 on GTX 1060 6GB GPU card driven by two i7 threads. Typically I got 7k-12k nodes per second, aside from some endgame positions, where the speed was much higher.
Houdini 4 was on one i7 3.8 GHz thread.
Time Control: 30 minutes base + 30 seconds increment, both engines.

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

Very interesting games, the chosen openings, and how Houdini 4 was pretty clueless for quite some time what is going on. One can say that only 2 games, sample is very small, but you can read here: viewtopic.php?t=63572 how these small samples can be interpreted.
It is quite likely that in LTC conditions, Lc0 NN329 with my domestic hardware is significantly above 3241 Elo rating of Houdini 4. And the scaling of Lc0 NN329 with time control is good (there were some problems with the scaling in the earlier nets).

[Lion]

I am clueless but on my 1060 gpu I have not more than 1kn/s.
How do you achieve 6=12 kn/s ?

Rgds

[Laskos]

I am clueless but on my 1060 gpu I have not more than 1kn/s.
How do you achieve 6=12 kn/s ?

Rgds

With these binaries
https://crem.xyz/lc0/
after installing NVIDIA CUDA and cuDNN packages.

[JJJ]

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 ?

[yanquis1972]

at LTC, maybe (but i doubt it)...at fast TC i really don't think lc0 could keep up with H4 on 1 core.

[Laskos]

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.

[George Tsavdaris]

From the standard opening position of Chess, I played two relatively long time control games, side and reversed, of LC0 and Houdini 4, a former top dog, and a 3241 CCRL 40/4' Elo rated engine on one thread.
Conditions are the following.
I used CUDA 9.2 version of GPU Lc0 with NN329 on GTX 1060 6GB GPU card driven by two i7 threads. Typically I got 7k-12k nodes per second, aside from some endgame positions, where the speed was much higher.
Houdini 4 was on one i7 3.8 GHz thread.
Time Control: 30 minutes base + 30 seconds increment, both engines.

It is quite likely that in LTC conditions, Lc0 NN329 with my domestic hardware is significantly above 3241 Elo rating of Houdini 4. And the scaling of Lc0 NN329 with time control is good (there were some problems with the scaling in the earlier nets).

From some matches from various sources i've seen, trying different than default settings for FPU reduction and CPUCT is giving a HUGE(not just big) boost on Lc0 in speed and more importantly strength. Setting FPU to -0.085 if i recall is one setting and CPUCT i don't remember. I saw default Lc0 lose with some -100 elo from one engine, and then with FPU and PUCT changed it won with some 30 elo!
Albert is posting some of these values also. So perhaps you can try these also.

[Laskos]

From the standard opening position of Chess, I played two relatively long time control games, side and reversed, of LC0 and Houdini 4, a former top dog, and a 3241 CCRL 40/4' Elo rated engine on one thread.
Conditions are the following.
I used CUDA 9.2 version of GPU Lc0 with NN329 on GTX 1060 6GB GPU card driven by two i7 threads. Typically I got 7k-12k nodes per second, aside from some endgame positions, where the speed was much higher.
Houdini 4 was on one i7 3.8 GHz thread.
Time Control: 30 minutes base + 30 seconds increment, both engines.

It is quite likely that in LTC conditions, Lc0 NN329 with my domestic hardware is significantly above 3241 Elo rating of Houdini 4. And the scaling of Lc0 NN329 with time control is good (there were some problems with the scaling in the earlier nets).

From some matches from various sources i've seen, trying different than default settings for FPU reduction and CPUCT is giving a HUGE(not just big) boost on Lc0 in speed and more importantly strength. Setting FPU to -0.085 if i recall is one setting and CPUCT i don't remember. I saw default Lc0 lose with some -100 elo from one engine, and then with FPU and PUCT changed it won with some 30 elo!
Albert is posting some of these values also. So perhaps you can try these also.

Yes, I used one of the latest CLOP results for these values:

NNCache size=200000
Scale thinking time=2.87
Cpuct MCTS Option=3.17
First Play Urgency Reduction=-0.068

I also left overnight play 2 games in the same condition from the standard opening position, Lc0 NN329 against Houdini 4 on 4 i7 threads (CCRL 40/4' rating of 3335), Lc0 won as White and drew as Black:
https://lichess.org/QrxXvZ8j
https://lichess.org/0ZXezfDC

And now the ultimate challenge: Lc0 NN329 as White against Stockfish 8 (1 thread, 3425 Elo) as Black, from standard opening position, one single game, but at double TC: 60 minutes base + 60 seconds increment. This game alone will take 4-6 hours. I am curious if Lc0 can hold a draw. Or even win?

[Nay Lin Tun]

Default LC0 setting from crem. 21 May, cuda 9.2 cuDNN version on 3GB, 1060 GTX, 2 cpu threads, average nodes/s for one minutes from initial positon= 7.6k average.
Nearly 7 times faster than the master Lczero(1.1knps), in my tests. Pretty amazing!!

[Laskos]

And now the ultimate challenge: Lc0 NN329 as White against Stockfish 8 (1 thread, 3425 Elo) as Black, from standard opening position, one single game, but at double TC: 60 minutes base + 60 seconds increment. This game alone will take 4-6 hours. I am curious if Lc0 can hold a draw. Or even win?

The game proceeded pretty uneventfully in French, Tarrasch, with Lc0 failing to gain any advantage in the opening and early midgame, where it usually builds its advantage. But it didn't blunder either, and throughout the game, the position was almost equal, and shown as such by both engines. Still, one LTC game and one draw against Stockfish 8 (1 thread), with Lc0 as White. The game is here:
https://lichess.org/07ZxMKrl

Sometime later, I will download the latest Stockfish dev, and put it on 4 threads in order to beat Lc0 NN329 playing White, at 3600''+ 60'' TC.