Twin 2080Tis and large NN

Discussion of anything and everything relating to chess playing software and machines.

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
shrapnel
Posts: 1221
Joined: Fri Nov 02, 2012 8:43 am
Location: New Delhi, India

Re: Twin 2080Tis and large NN

Post by shrapnel » Sat Oct 26, 2019 5:24 am

corres wrote:
Fri Oct 25, 2019 7:32 pm
Opening positions was the same in both matches.
Leela parameters were also the same in the case of the two net.
So enhancing the move time (or the power of GPUs) the result tend to reach equality.
It would be interesting to make a match with Time = 1 min + 200 sec / move (~ FIDE tempo)
This is the mistake the Testers at lc0.org are making, at least as far as Competitive Chess against AB Engines is concerned.
While your Tests are very interesting and have their own significance, there is a BIG difference between playing different NNs of lc0 against each other and playing lc0 against powerful AB Engines like Stockfish and its derivatives.
For example, if I use one of the smaller NNs like the 128x10 or 256x20 on my System, lc0 gets "smoked" nearly every time against even relatively weak CPUs running Stockfish, no matter the very high Knps or Time Control used.
Its ONLY using the very large 384x30 Net that enables lc0 to match upto the AB Engine, whatever the TC. Of course, a strong GPU or two is required to fully utilize the large NN, unless one wants to run it at 1 Knps or so, which is again pointless.
So, in the Real World (in a manner of speaking) of competitive Engine-Engine Matches a very large NN is required, supported by a strong GPU or two, at least at the present stage of lc0 development, no matter what the Theorists say.
i7 5960X @ 4.1 Ghz, 64 GB G.Skill RipJaws RAM, Twin Asus ROG Strix OC 11 GB Geforce 2080 Tis

corres
Posts: 1681
Joined: Wed Nov 18, 2015 10:41 am
Location: hungary

Re: Twin 2080Tis and large NN

Post by corres » Sat Oct 26, 2019 11:10 am

shrapnel wrote:
Sat Oct 26, 2019 5:24 am
corres wrote:
Fri Oct 25, 2019 7:32 pm
Opening positions was the same in both matches.
Leela parameters were also the same in the case of the two net.
So enhancing the move time (or the power of GPUs) the result tend to reach equality.
It would be interesting to make a match with Time = 1 min + 200 sec / move (~ FIDE tempo)
This is the mistake the Testers at lc0.org are making, at least as far as Competitive Chess against AB Engines is concerned.
While your Tests are very interesting and have their own significance, there is a BIG difference between playing different NNs of lc0 against each other and playing lc0 against powerful AB Engines like Stockfish and its derivatives.
For example, if I use one of the smaller NNs like the 128x10 or 256x20 on my System, lc0 gets "smoked" nearly every time against even relatively weak CPUs running Stockfish, no matter the very high Knps or Time Control used.
I think if you would use a special opening book what is appropriate to the stile of NN engines and it utilizes the weakness of AB engines you could get better results. Maybe the "Attacker Book" is not the best for an NN engine.
shrapnel wrote:
Sat Oct 26, 2019 5:24 am
Its ONLY using the very large 384x30 Net that enables lc0 to match upto the AB Engine, whatever the TC. Of course, a strong GPU or two is required to fully utilize the large NN, unless one wants to run it at 1 Knps or so, which is again pointless.
So, in the Real World (in a manner of speaking) of competitive Engine-Engine Matches a very large NN is required, supported by a strong GPU or two, at least at the present stage of lc0 development, no matter what the Theorists say.
In general a good big NN is stronger than a small one and it needs one or more stronger GPU(s)
to realize its full potential.
But TCEC and CCC proved there are strong mid range 256x20 nets also. It is pity they need
3-4 x RTX 2080 TI against Stockfish running on powerful PC.
Btw.
What a kind of engine is "Eman"?
As I know it is a weaker derivative of Stockfish - like Sugar, CorChess, etc.

Leo
Posts: 839
Joined: Fri Sep 16, 2016 4:55 pm
Location: USA/Minnesota
Full name: Leo

Re: Twin 2080Tis and large NN

Post by Leo » Sat Oct 26, 2019 7:24 pm

mwyoung wrote:
Fri Oct 25, 2019 5:20 pm
shrapnel wrote:
Fri Oct 25, 2019 4:11 pm
Large NNs (384x30) really do improve the performance of lc0( whether Laskos Sir fully agrees or not ). Of course, twin 2080 Tis help :) , but on the other hand using a large NN with a weak GPU makes no sense and would simply backfire.
Following is a Game against a guy using 16 Cores and Stockfish..



Against a slightly weaker opponent, lc0 won, even though it was a 3 minute Game, supposedly not lc0's forte....

Thanks for the games. It looks like the best CPU you played was a older 8 core 2700x CPU, and this game was a draw. I have not seen any data that suggest using multi GPUs with Lc0 is worth the investment at this point. But this could change.... Right now SF is just smoking Lc0 and the other NN in my testing.

Stockfish and a good CPU provides the best chess value and strength.
Do you know why SF lost to LCO in the superfinal 2 seasons ago? How did it win the latest TCEC?
Advanced Micro Devices fan.

shrapnel
Posts: 1221
Joined: Fri Nov 02, 2012 8:43 am
Location: New Delhi, India

Re: Twin 2080Tis and large NN

Post by shrapnel » Sun Oct 27, 2019 3:10 am

corres wrote:
Sat Oct 26, 2019 11:10 am
Maybe the "Attacker Book" is not the best for an NN engine.
Agreed. Can you point out a Book more suitable for lc0 ?
In fact, I don't like using an Opening Book for lc0, but unfortunately the AB Engines slaughter it if it doesn't use one.
So, no choice.
corres wrote:
Sat Oct 26, 2019 11:10 am
What kind of engine is "Eman"?
I don't know if its weaker than Stockfish, but it's a very popular Engine being used many Players on Playchess and InfinityChess, so it must be quite good.
i7 5960X @ 4.1 Ghz, 64 GB G.Skill RipJaws RAM, Twin Asus ROG Strix OC 11 GB Geforce 2080 Tis

corres
Posts: 1681
Joined: Wed Nov 18, 2015 10:41 am
Location: hungary

Re: Twin 2080Tis and large NN

Post by corres » Sun Oct 27, 2019 10:26 am

shrapnel wrote:
Sun Oct 27, 2019 3:10 am
corres wrote:
Sat Oct 26, 2019 11:10 am
Maybe the "Attacker Book" is not the best for an NN engine.
Agreed. Can you point out a Book more suitable for lc0 ?
In fact, I don't like using an Opening Book for lc0, but unfortunately the AB Engines slaughter it if it doesn't use one.
So, no choice.
[/quote]

I also do not know special book for NN engines. But I think because of the similarity between playing stile of humans and NN engines a book made from well selected GM games may be appropriate to NN engines too. Another opportunity: You collect the games of NN engines (won and drawn) from TCEC and CCC and other site and you make from them a book.
In general it is a good strategy to use closed openings to NN engines. Stockfis is too strong in opened and half opened games.

corres
Posts: 1681
Joined: Wed Nov 18, 2015 10:41 am
Location: hungary

Re: Twin 2080Tis and large NN

Post by corres » Sun Oct 27, 2019 10:27 am

shrapnel wrote:
Sun Oct 27, 2019 3:10 am
corres wrote:
Sat Oct 26, 2019 11:10 am
Maybe the "Attacker Book" is not the best for an NN engine.
Agreed. Can you point out a Book more suitable for lc0 ?
In fact, I don't like using an Opening Book for lc0, but unfortunately the AB Engines slaughter it if it doesn't use one.
So, no choice.
I also do not know special book for NN engines. But I think because of the similarity between playing stile of humans and NN engines a book made from well selected GM games maybe appropriate to NN engines too. Another opportunity: You collect the games of NN engines (won and drawn) from TCEC and CCC and other site and you make from them a book.
In general it is a good strategy to use closed openings with NN engines. Stockfis is too strong in opened and half opened games.

Hai
Posts: 455
Joined: Sun Aug 04, 2013 11:19 am

Re: Twin 2080Tis and large NN

Post by Hai » Tue Oct 29, 2019 8:22 am

shrapnel wrote:
Fri Oct 25, 2019 4:11 pm
Large NNs (384x30) really do improve the performance of lc0( whether Laskos Sir fully agrees or not ). Of course, twin 2080 Tis help :) , but on the other hand using a large NN with a weak GPU makes no sense and would simply backfire.
That's right.
I'm using 30x384 and 2x RTX 2080 Ti and it's great:)
LC0 is playing much stronger and finds more key-moves, than before. The strength is increasing much more with more time.

User avatar
Nordlandia
Posts: 2457
Joined: Fri Sep 25, 2015 7:38 pm
Location: Sortland, Norway

Re: Twin 2080Tis and large NN

Post by Nordlandia » Tue Oct 29, 2019 9:22 am

On paper 2080 Ti SLi is a monster.

shrapnel
Posts: 1221
Joined: Fri Nov 02, 2012 8:43 am
Location: New Delhi, India

Re: Twin 2080Tis and large NN

Post by shrapnel » Tue Oct 29, 2019 3:18 pm

Hai wrote:
Tue Oct 29, 2019 8:22 am
That's right.
I'm using 30x384 and 2x RTX 2080 Ti and it's great:)
LC0 is playing much stronger and finds more key-moves, than before. The strength is increasing much more with more time.
Yes, indeed. The Sergei 30x384 nets are better than anything that the official lco.org has produced. They're simply wasting time and resources on 128x24 nets, though I've been told that they are simply a means to an end and ultimately very strong nets will be produced.
In the meantime, Sergei's NNs will do very nicely for me :) .
Nordlandia wrote:
Tue Oct 29, 2019 9:22 am
On paper 2080 Ti SLi is a monster.
On paper ? Get a load of this Game against a 36 Core Xeon Monster using Stockfish
i7 5960X @ 4.1 Ghz, 64 GB G.Skill RipJaws RAM, Twin Asus ROG Strix OC 11 GB Geforce 2080 Tis

shrapnel
Posts: 1221
Joined: Fri Nov 02, 2012 8:43 am
Location: New Delhi, India

Re: Twin 2080Tis and large NN

Post by shrapnel » Tue Oct 29, 2019 5:27 pm

Oh, here's another, where my bada$$ 2080 Tis gang up to beat the Brains out of the latest, much-vaunted Brainfish....BRUTAL... :)
i7 5960X @ 4.1 Ghz, 64 GB G.Skill RipJaws RAM, Twin Asus ROG Strix OC 11 GB Geforce 2080 Tis

Post Reply