Page 1 of 2
Big boost in strength for LC0 by simply changing VGA card
Posted: Fri Jan 25, 2019 8:08 am
by M ANSARI
I have been playing around with LC0 on a pretty good VGA card but to be honest, but in quick tourneys it was getting trounced by even SF7 with SF using 6 core Intel CPU. So I decided to go for a good VGA card and got a RTX 2080 Ti. Man was that thing huge when it came in !!! I had to move all my computer innards to a larger case to allow for the length of the card. It also weighed more that my full sized laptop. Anyway I was expecting some improvement, but was pretty surprised that the boost in strength was quite substantial. I did a quick tournament of 20 games against one of the latest Dev SF and Leela I used a Network that was around 3600. Lc0 really smashed SF and looking at some games quickly, it seemed that the score could have been even higher as Lc0 squandered some winning positions as it started playing on increment. I could see only one loss for Lc0 where it allowed a queen deep against its defenseless King, but otherwise it just simply seemed to outplay SF every game. Maybe there is a hardware disparity with a RTX 2080 Ti and 6 Cpu @4 Ghz for SF. I have to admit the chess it plays is very hard to comprehend as it completely ignores some things in chess that we are taught are bad and maybe at 3_2 and such a short set of games this is nothing. But this Lc0 stuff is certainly not hype and there is no doubt in my mind now that this is a turning point in how chess engines are made. I mean it is still only 10 months out since all this started and it already is a beast. There certainly is a huge potential for improvement as everything is still experimental. I can't imagine how things are once a mature network is found and once drivers and better utilization of available hardware is there.
LC0_SFDev 061119, Blitz 3m+2s 0
1 Lc0 v0.20.1 +89 +6/=13/-1 62.50% 12.5/20
2 Stockfish 060119 64 POPCNT -89 +1/=13/-6 37.50% 7.5/20
How do you add the pgn files of the games? Somehow I get an error if I try to attach the pgn file of the games. Can I just add it as a .cbh file?
Re: Big boost in strength for LC0 by simply changing VGA card
Posted: Fri Jan 25, 2019 8:45 am
by Guenther
M ANSARI wrote: ↑Fri Jan 25, 2019 8:08 am
...
LC0_SFDev 061119, Blitz 3m+2s 0
1 Lc0 v0.20.1 +89 +6/=13/-1 62.50% 12.5/20
2 Stockfish 060119 64 POPCNT -89 +1/=13/-6 37.50% 7.5/20
How do you add the pgn files of the games? Somehow I get an error if I try to attach the pgn file of the games. Can I just add it as a .cbh file?
Why don't you just add it as a zipped/7z file?
Plain pgn is not supported as attachment (just archive and image formats) in the shown list of file extensions (too big!).
Re: Big boost in strength for LC0 by simply changing VGA card
Posted: Fri Jan 25, 2019 9:46 am
by M ANSARI
Ok I can do that. Just seems strange that a chess forum wouldn't support some sort of chess database format rather than to have to zip the files.
Re: Big boost in strength for LC0 by simply changing VGA card
Posted: Fri Jan 25, 2019 10:24 am
by jmartus
Thats wild 90 elo difference stockfish is like two years behind lco
Re: Big boost in strength for LC0 by simply changing VGA card
Posted: Fri Jan 25, 2019 11:08 pm
by mclane
M ANSARI wrote: ↑Fri Jan 25, 2019 8:08 am
I have been playing around with LC0 on a pretty good VGA card but to be honest, but in quick tourneys it was getting trounced by even SF7 with SF using 6 core Intel CPU. So I decided to go for a good VGA card and got a RTX 2080 Ti. Man was that thing huge when it came in !!! I had to move all my computer innards to a larger case to allow for the length of the card. It also weighed more that my full sized laptop. Anyway I was expecting some improvement, but was pretty surprised that the boost in strength was quite substantial. I did a quick tournament of 20 games against one of the latest Dev SF and Leela I used a Network that was around 3600. Lc0 really smashed SF and looking at some games quickly, it seemed that the score could have been even higher as Lc0 squandered some winning positions as it started playing on increment. I could see only one loss for Lc0 where it allowed a queen deep against its defenseless King, but otherwise it just simply seemed to outplay SF every game. Maybe there is a hardware disparity with a RTX 2080 Ti and 6 Cpu @4 Ghz for SF. I have to admit the chess it plays is very hard to comprehend as it completely ignores some things in chess that we are taught are bad and maybe at 3_2 and such a short set of games this is nothing. But this Lc0 stuff is certainly not hype and there is no doubt in my mind now that this is a turning point in how chess engines are made. I mean it is still only 10 months out since all this started and it already is a beast. There certainly is a huge potential for improvement as everything is still experimental. I can't imagine how things are once a mature network is found and once drivers and better utilization of available hardware is there.
LC0_SFDev 061119, Blitz 3m+2s 0
1 Lc0 v0.20.1 +89 +6/=13/-1 62.50% 12.5/20
2 Stockfish 060119 64 POPCNT -89 +1/=13/-6 37.50% 7.5/20
How do you add the pgn files of the games? Somehow I get an error if I try to attach the pgn file of the games. Can I just add it as a .cbh file?
You change the hardware and wonder that the outcome is different ?
Re: Big boost in strength for LC0 by simply changing VGA card
Posted: Sat Jan 26, 2019 3:06 am
by jmartus
TCEC finals will be very intresting!
Re: Big boost in strength for LC0 by simply changing VGA card
Posted: Sat Jan 26, 2019 3:17 am
by dkappe
$350-$1000 gets you hardware that completely blows ab engines on commodity hardware out of the water. Welcome to the leela club.
Re: Big boost in strength for LC0 by simply changing VGA card
Posted: Sat Jan 26, 2019 3:18 am
by Geonerd
What card were you running before? What NPS did it generate?
Have you enabled the FP16 back end on the new card? What NPS are you getting now?
Re: Big boost in strength for LC0 by simply changing VGA card
Posted: Sat Jan 26, 2019 7:55 am
by M ANSARI
Card I am using is now an RTX 2080Ti which reading on the forums seems to indicate would do very well. The previous card was a Quadro M4000 with around 1700 Cuda cores while the RTX 2080Ti has about 3.5x that. Of course there must be some Mhz speed differences and probably better memory latency performance.
I am doing another match now, but this time using a Silvesuite book and things are not going as well. At the moment after 60 games Leela is down by 2 games with score at +11, =36,-13. I looked at some of the openings from book and many openings had SF up by over +1 from the opening and Leela on many openings was very pessimistic about the position it had to start with. Most likely I will need a more neutral book. I wanted to try out the Nunn short book but couldn't find it somehow. Still, on the same system (except for the GPU) LC0 was not doing well even against SF7 ... so obviously GPU hardware plays a big factor. I couldn't help but notice that LC0 chess play seems very raw with wildly fluctuating evaluations and somehow it seems to take the long route for wins. So you could say the engine is still immature in its play, but I think that once that is sorted out there will be a huge gap between AB engines and this type of engine.
I have to admit all this stuff has me close to biting the bug. I am almost tempted to start a crazy build with my now dormant custom double vapor chilller that I build many years back. Many years back I had it running a dual socket Skulltrail with 8 cores at 5 Ghz and it rocked for chess. That was some crazy speed back then. I need to check if things can be overclocked on the GPU end or maybe a better option is just to add multiple cards. I do remember that having 8 cores at 5Ghz played much stronger than 16 cores at 2.8 Ghz ... so maybe something like that is possible with GPU cards. It might be an interesting project.
Re: Big boost in strength for LC0 by simply changing VGA card
Posted: Sat Jan 26, 2019 11:44 am
by Sesse
The RTX 2000 series has specialized hardware for neural networks. These “tensor cores” is the reason why your new GPU blows the old one out of the water—the number of CUDA cores simply doesn't matter, as they're not the ones doing the work.