game showing some blindness by Lc0

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.
Post Reply
jdart
Posts: 3879
Joined: Fri Mar 10, 2006 4:23 am
Location: http://www.arasanchess.org

game showing some blindness by Lc0

Post by jdart » Sun Oct 13, 2019 10:14 pm

This was played on playchess.com today. White (Lc0) seems to have made a couple serious errors here. One is 31. Rxh6?:



White took the wrong pawn: it was much better to capture the dangerous passer with Qxd3. Also note 36. Rh7? Better was Kg1, a move Stockfish and Houdini find quickly:



After this Lc0 had negative evals for a while but they were quite moderate compared to its opponent. It did not seem to see the danger. In the final endgame material is even but Black is winning.


Nay Lin Tun
Posts: 552
Joined: Mon Jan 16, 2012 5:34 am

Re: game showing some blindness by Lc0

Post by Nay Lin Tun » Mon Oct 14, 2019 4:43 am

Do you know the net version , GPU and approx speed of Lco?

jdart
Posts: 3879
Joined: Fri Mar 10, 2006 4:23 am
Location: http://www.arasanchess.org

Re: game showing some blindness by Lc0

Post by jdart » Mon Oct 14, 2019 10:40 am

I do not know. It is a server account. Rating on playchess is 2397.

--Jon

shrapnel
Posts: 1261
Joined: Fri Nov 02, 2012 8:43 am
Location: New Delhi, India

Re: game showing some blindness by Lc0

Post by shrapnel » Mon Oct 14, 2019 7:44 pm

jdart wrote:
Sun Oct 13, 2019 10:14 pm
This was played on playchess.com today. White (Lc0) seems to have made a couple serious errors here. One is 31. Rxh6?:

Blindness, my FOOT !
lc0 will play only as strong as the Hardware and Neural Nets it gets, unlike the AB Engines using which, even a relatively weak 4-Core Player can hold a 16 Core player to a Draw, quite often.
Using my new 2080 Ti which just arrived and am still Tuning and the very powerful and large Sergio NN which is 384x30 ( which even slows down the mighty 2080 Ti) lc0 found Qxd3 in just over 30 seconds.
Image
This just confirms my view that people with weak GPUs shouldn't even bother trying out lco, as it just gives a bad reputation to the Engine when they get beaten badly.
lc0 NEEDS strong GPUs and large NNs to realize its full potential. Smaller Networks give more Knp/s no doubt, but they are ultimately too weak to hold their own against Stockfish/Stockfish derivative players.
In fact I think the folk over at lczero.org are wasting their time doing Training runs using the very weak 128x10 NNs.
Despite the very high Knps I get, I usually get beaten or draw on Infinity Chess.
The 320x24 NNs are much better and I rarely lose.
The GPU should be able to run the cudnn-fp-16 backend at the very least, if it wants to be taken seriously.
i7 5960X @ 4.1 Ghz, 64 GB G.Skill RipJaws RAM, Twin Asus ROG Strix OC 11 GB Geforce 2080 Tis

jdart
Posts: 3879
Joined: Fri Mar 10, 2006 4:23 am
Location: http://www.arasanchess.org

Re: game showing some blindness by Lc0

Post by jdart » Tue Oct 15, 2019 1:33 am

This was a rapid game so White may not have had 30 seconds to make the right move, and would have required even more than that if running on lesser hardware. Stockfish will find Qxd3 right away.

glennsamuel32
Posts: 89
Joined: Sat Dec 04, 2010 4:31 pm
Location: 223

Re: game showing some blindness by Lc0

Post by glennsamuel32 » Tue Oct 15, 2019 3:43 am

Using my new 2080 Ti which just arrived and am still Tuning and the very powerful and large Sergio NN which is 384x30 ( which even slows down the mighty 2080 Ti) lc0 found Qxd3 in just over 30 seconds.
Lc0 finds a move when that move is the "first focus" -- your pic shows it still hasn't chosen it.
My laptop finds it in 64 secs...

Image

shrapnel
Posts: 1261
Joined: Fri Nov 02, 2012 8:43 am
Location: New Delhi, India

Re: game showing some blindness by Lc0

Post by shrapnel » Tue Oct 15, 2019 6:23 am

jdart wrote:
Tue Oct 15, 2019 1:33 am
This was a rapid game so White may not have had 30 seconds to make the right move, and would have required even more than that if running on lesser hardware. Stockfish will find Qxd3 right away.
True. Rapid games are certainly not the forte of GPU-intensive Engines.Even the first AlphaZero-Stockfish Match was played at 1 min/move. So the Google Brains were also well aware of this.
All the more reason that people with weak GPUs shouldn't even look at lc0.
On the positive side, GPUs are only getting stronger and stronger and lc0 performance can only improve, provided you have the money.
Thomas Spark over at Chessbase Server is running TWO 2080 Tis and is rarely beaten and has a ELO Rating of 2548.
A GPU faster than the 2080 Ti is being released in March '20 and the Networks will only get stronger.
So plenty to look forward to....
i7 5960X @ 4.1 Ghz, 64 GB G.Skill RipJaws RAM, Twin Asus ROG Strix OC 11 GB Geforce 2080 Tis

shrapnel
Posts: 1261
Joined: Fri Nov 02, 2012 8:43 am
Location: New Delhi, India

Re: game showing some blindness by Lc0

Post by shrapnel » Tue Oct 15, 2019 6:29 am

glennsamuel32 wrote:
Tue Oct 15, 2019 3:43 am
Lc0 finds a move when that move is the "first focus" -- your pic shows it still hasn't chosen it.
It does, just a few seconds later. I was just trying to highlight the earliest when it started considering that move seriously.
With TWO 2080 Tis, like what Thomas Spark over at the Chessbase Server has, the time taken will become almost half.
I too, am considering a 2nd 2080 Ti or I may wait for the new nVidia GPU which is being released in the first quarter of 2020.
Cheers !
i7 5960X @ 4.1 Ghz, 64 GB G.Skill RipJaws RAM, Twin Asus ROG Strix OC 11 GB Geforce 2080 Tis

Post Reply