Big boost in strength for LC0 by simply changing VGA card

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

Moderators: hgm, Harvey Williamson, bob

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
brianr
Posts: 356
Joined: Thu Mar 09, 2006 2:01 pm

Re: Big boost in strength for LC0 by simply changing VGA card

Post by brianr » Sat Jan 26, 2019 11:37 am

M ANSARI wrote:
Sat Jan 26, 2019 6:55 am
...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.
Books are actually a handicap for Leela. The Silver and Nunn books are great, just not as helpful with Leela as with "traditional" alpha/beta engines. Suggest using only a two ply (1 move each side) only book depth.

ankan
Posts: 71
Joined: Sun Apr 21, 2013 1:29 pm
Full name: Ankan Banerjee
Contact:

Re: Big boost in strength for LC0 by simply changing VGA card

Post by ankan » Sat Jan 26, 2019 11:42 am

M ANSARI wrote:
Sat Jan 26, 2019 6:55 am
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.
I would expect a bigger jump in NPS going from Quadro M4000 to the RTX 2080Ti (something like 10x or even more).
Make sure you have --backend=cudnn-fp16 as command line option (or select it from GUI's engine settings). lc0 utilizes tensor cores only with the cudnn-fp16 backend.
Increasing the NNCache size (from the default 200k to something like 10 million) can also boost NPS a bit without any side-effects (add --nncache=10000000 to command line settings or in the GUI's engine settings).

User avatar
M ANSARI
Posts: 3407
Joined: Thu Mar 16, 2006 6:10 pm

Re: Big boost in strength for LC0 by simply changing VGA card

Post by M ANSARI » Sat Jan 26, 2019 2:31 pm

ankan wrote:
Sat Jan 26, 2019 11:42 am
M ANSARI wrote:
Sat Jan 26, 2019 6:55 am
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.
I would expect a bigger jump in NPS going from Quadro M4000 to the RTX 2080Ti (something like 10x or even more).
Make sure you have --backend=cudnn-fp16 as command line option (or select it from GUI's engine settings). lc0 utilizes tensor cores only with the cudnn-fp16 backend.
Increasing the NNCache size (from the default 200k to something like 10 million) can also boost NPS a bit without any side-effects (add --nncache=10000000 to command line settings or in the GUI's engine settings).
Thanks for this advice. I did not have those settings so I am guessing this will improve things. I now get 52kn/s on the start position (not sure what it was before the setting change) using the network 32591. I am guessing there is probably some better networks out there ... but really everything is so much in flux right now that I feel things will be changing a lot for while. I was reading up on Tensor cores and it makes for some interesting reading. Apparently this is all done for AI with Google having been the starter of this but several other companies are starting to make even bigger headway in that field (nvidia being one of them). My guess is these Tensor cores are similar what the Pentium Pro a little over 2 decades ago was then. The hardware and software to use the hardware will probably see an exponential improvement as it becomes more mainstream. If LC0 is so strong already, I can't imagine what things will be like in a few years. There are some pretty huge holes in LC0 evaluation ... in one game I saw that it just did not realize how strong it was for the other side to gain 2 queens ...somehow it just gave it almost 1 pawn advantage for the extra queen while SF was at +10 and obviously SF was correct. These things will be patched up soon I think. I remember not too long ago Rybka in its prime, on my 8 core 5 Ghz system was crushing everything on Playchess ... but then in one game it completely missed a wrong bishop endgame and went head first into it thinking it was winning while it was a draw. Those type of things need to be patched up and if the rate of progress continues, it will only be a matter of time.

jmartus
Posts: 198
Joined: Sun May 16, 2010 12:50 am

Re: Big boost in strength for LC0 by simply changing VGA card

Post by jmartus » Sat Jan 26, 2019 3:09 pm

Net 32335 is the strongest one. Should try it with that one.

carldaman
Posts: 1706
Joined: Sat Jun 02, 2012 12:13 am

Re: Big boost in strength for LC0 by simply changing VGA card

Post by carldaman » Sat Jan 26, 2019 9:34 pm

jmartus wrote:
Sat Jan 26, 2019 3:09 pm
Net 32335 is the strongest one. Should try it with that one.
If that's the case, why did they pick the later 32742 for TCEC?

It is very hard to get non-contradictory information about LCO these days.

jmartus
Posts: 198
Joined: Sun May 16, 2010 12:50 am

Re: Big boost in strength for LC0 by simply changing VGA card

Post by jmartus » Sat Jan 26, 2019 11:17 pm

I'm not 100 percent sure i was just going off someones ccrl rating list in the discord forum.

Nay Lin Tun
Posts: 514
Joined: Mon Jan 16, 2012 5:34 am

Re: Big boost in strength for LC0 by simply changing VGA card

Post by Nay Lin Tun » Sun Jan 27, 2019 2:11 am

jmartus wrote:
Sat Jan 26, 2019 11:17 pm
I'm not 100 percent sure i was just going off someones ccrl rating list in the discord forum.
ccrl rating list by MGTO is pretty incorrect indicator of NN rating for LTC or strong hardware like TCEC/CCcc. Lco is practically a hybrid engine and the strength of engine is the result of a combination of NN's policy strength(similar to static evaluation of Stockfish at 0 node) and value head strength( MCTS search and its evaluation which is similar to Stockfish evaluation at depth 30~50 etc ).
When bullet with 4k nps like ccrl list, the value of policy head matters cos the search is very limited.
But it Tcec/cccc, they have 60~ 100knps and also VLTC, and the better value head matters more. According to test results of big hardwares, 32700+ or 32800+ outperform the network u mentioned above.

yanquis1972
Posts: 1762
Joined: Tue Jun 02, 2009 10:14 pm

Re: Big boost in strength for LC0 by simply changing VGA card

Post by yanquis1972 » Sun Jan 27, 2019 3:21 am

suspect there's been a significant gain in conjunction with the drop in self-elo, which seems to've recovered. a lot of complex endgames seem soft-solved; winning scores (~75%/~+2.50) where SF might give half/three quarters pawn advantage with fairly deep search.

this 1min+1s bullet game on weak hardware (~2.5MN/s SF10; ~6kn/s Lc0) deserves some much deeper analysis but is a complete spectacle without it -- https://lichess.org/fAEh9NgM




User avatar
M ANSARI
Posts: 3407
Joined: Thu Mar 16, 2006 6:10 pm

Re: Big boost in strength for LC0 by simply changing VGA card

Post by M ANSARI » Sun Jan 27, 2019 6:34 am

I did another 3_2 game, this time 50 games with the new settings and it seems score is very similar to the initial 20 game set. Lc0 is at +16 -6 and the rest draws after 40 games. This is around +90 ELO which is quite impressive. Again this is with no books. I have to say I am really impressed. There was one game where Lc0 crazily gave up its queen and went into an ending thinking it was equal. I mean I guarantee that 100% of all super GM's and every engine would disagree as it seemed there was no fortress in the endgame. SF had the evaluation at around +2.4 ... but it turns out that Lc0 knew that it could force off one pair of rooks and then get a fortress. I was following that game live and I thought that either Lc0 was a genius and or was going to go down in flames ... I guess after about 40 moves SF seemed to agree there was no progress so Lc0 was probably correct. For me that was mind boggling as the weakness of engines was that they could not understand fortresses with inferior material or had a blindness for the horizon effect. Not anymore I guess! Once the game is over I will try to filter that game out as I thought it was game changing.

Javier Ros
Posts: 181
Joined: Fri Oct 12, 2012 10:48 am
Location: Seville (SPAIN)
Full name: Javier Ros

Re: Big boost in strength for LC0 by simply changing VGA card

Post by Javier Ros » Sun Jan 27, 2019 6:48 pm

M ANSARI wrote:
Sun Jan 27, 2019 6:34 am
I did another 3_2 game, this time 50 games with the new settings and it seems score is very similar to the initial 20 game set. Lc0 is at +16 -6 and the rest draws after 40 games. This is around +90 ELO which is quite impressive. Again this is with no books. I have to say I am really impressed. There was one game where Lc0 crazily gave up its queen and went into an ending thinking it was equal. I mean I guarantee that 100% of all super GM's and every engine would disagree as it seemed there was no fortress in the endgame. SF had the evaluation at around +2.4 ... but it turns out that Lc0 knew that it could force off one pair of rooks and then get a fortress. I was following that game live and I thought that either Lc0 was a genius and or was going to go down in flames ... I guess after about 40 moves SF seemed to agree there was no progress so Lc0 was probably correct. For me that was mind boggling as the weakness of engines was that they could not understand fortresses with inferior material or had a blindness for the horizon effect. Not anymore I guess! Once the game is over I will try to filter that game out as I thought it was game changing.

I had the same impressions when I played the first games with Lc0. I am still amazed when I see how it is able to turn a game where Stockfish evaluates with advantage sacrificing a pair of pawns. The following game was played between SF10 on 6 cores of i7 980X and AntiFish125 on 2x1060 played at 5 minutes plus 1 second.

After eating the second pawn SF10 evaluates as +0.86

Image

but the advanced black pawns of d and e worsen the coordination of the white pieces and the threat of the column g gives black advantage. After 29.g3 +0.50 if follows

Image

The love relationship between a chess engine tester and his computer can be summarized in one sentence:
Until heat do us part.

Post Reply