Latest Updates:
Page Index Toggle Pages: [1] 2 3 ... 5
Topic Tools
Very Hot Topic (More than 25 Replies) New version of LcO (Read 7910 times)
Jupp53
God Member
*****
Offline


be

Posts: 873
Location: Frankfurt/Main
Joined: 01/04/09
Gender: Male
Re: New version of LcO
Reply #72 - 11/19/19 at 14:21:22
Post Tools
The german newspaper ct' has in its current edition (24/2019; till this weekend) a proposal for a silent pc fitting to these ideas. I think if you don't want to play competitive correspondence chess on highest level one RTX 2070 or 2060 model are sufficient for having good support in working for yourself.

The pc is cheaper with two GPUs than my first one 36 years ago and I had less money than. So power efficiency for the climate is my main concern meanwhile.
  

Medical textbooks say I should be dead since April 2002.
Dum spiro spero. Smiley
Back to top
 
IP Logged
 
JFugre
Junior Member
**
Offline


I Love ChessPublishing!

Posts: 74
Joined: 01/22/19
Re: New version of LcO
Reply #71 - 11/18/19 at 23:11:50
Post Tools
IsaVulpes wrote on 11/18/19 at 17:10:13:
RTX2060 - 18500 NPS - $320
GTX1660ti - 8500 NPS - $260
+117% performance, for +23% price


The GTX 1660 Ti is incredibly cost inefficient, no matter the application (Leela or games). In Europe, it seems to cost almost as much as the RTX 2060.

The non-Ti version is 40% cheaper. But he was looking for a laptop so the options are always rather different then.

The important thing is to not get a non-Turing card (i.e. GTX 10x0 series), or you'll pay more for ~25% less performance. RTX is nice if you can afford it.

The ChessBase guide for system building is hilarious. Building a system with 2 x RTX 2070 and a 2700X CPU? It will be hot, power hungry and a nightmare to keep quiet. Useful for showing off on playchess or in correspondence, not so useful if you actually study for OTB.
  
Back to top
 
IP Logged
 
IsaVulpes
Senior Member
****
Offline


No.

Posts: 289
Joined: 12/09/07
Re: New version of LcO
Reply #70 - 11/18/19 at 17:10:13
Post Tools
IsaVulpes wrote on 11/09/19 at 23:15:56:
General benchmarks are of very little use for a comparison between RTX cards and others for Leela purposes

The RTX cards are boosted with tensor cores, which are an incredible performance boost for all machine learning applications.

RTX 2060 and GTX 1660i are at a similar pricepoint, but the former should bring out vastly more power for Leela (whether anyone needs that, is of course a different question).
RTX 2080 will outperfom any GTX card by miles.


https://en.chessbase.com/post/fat-fritz-what-videocard-to-buy Here is a chessbase article underlining what I said

RTX2060 - 18500 NPS - $320
GTX1660ti - 8500 NPS - $260
+117% performance, for +23% price

If you want to buy a GPU for Leela, get an NVidia RTX card.
  
Back to top
 
IP Logged
 
IsaVulpes
Senior Member
****
Offline


No.

Posts: 289
Joined: 12/09/07
Re: New version of LcO
Reply #69 - 11/09/19 at 23:15:56
Post Tools
General benchmarks are of very little use for a comparison between RTX cards and others for Leela purposes

The RTX cards are boosted with tensor cores, which are an incredible performance boost for all machine learning applications.

RTX 2060 and GTX 1660i are at a similar pricepoint, but the former should bring out vastly more power for Leela (whether anyone needs that, is of course a different question).
RTX 2080 will outperfom any GTX card by miles.
  
Back to top
 
IP Logged
 
brabo
God Member
*****
Offline


Welcome chessfriend

Posts: 917
Joined: 02/02/07
Re: New version of LcO
Reply #68 - 10/21/19 at 08:35:33
Post Tools
I finally bought a new laptop:  https://www.alternate.be/ASUS/TUF-Gaming-FX505DU-AL057T-BE/html/product/1555008?...

I estimate that the Nvidea GTX 1660i is about 40 elo weaker than the recommended Nvidea RTX 2080 see https://gpu.userbenchmark.com/Compare/Nvidia-RTX-2080-vs-Nvidia-GTX-1660-Ti/4026...

However I didn't want to spend 2-3 times more money for such small gain.

I compared the new portable with my old desktop (My AMD FX(tm)-6300 6 core processor 3,5Ghz is configured with a Nvidia Geforce GTX 960.) and noted the following gains of speed:
Komodo 11 and Stockfish 10: + 70%
Lc0: + 1000% (using Cuda instead of opencl)

I organized 2 new 100 rapidmatches (Lc0 used net 42850)
Lc0 - Komodo 11: 62,5 - 37,5
Lc0 - Stockfish 10: 52,5 - 47,5

Below just one amazing game from Lc0 against Stockfish:

The return-game was easily drawn by Lc0.
  
Back to top
WWW  
IP Logged
 
fling
God Member
*****
Offline


I Love ChessPublishing!

Posts: 1550
Joined: 01/21/11
Gender: Male
Re: New version of LcO
Reply #67 - 09/20/19 at 17:37:44
Post Tools
Ok. Sorry for commenting. Good luck in the future!
  
Back to top
 
IP Logged
 
brabo
God Member
*****
Offline


Welcome chessfriend

Posts: 917
Joined: 02/02/07
Re: New version of LcO
Reply #66 - 09/20/19 at 05:56:43
Post Tools
The questions are all answered in my article.

The matches with Lc0 were all organized in the months May, June and July with the release which was at that time available (see dates of the games). I know around that period they were bringing out new engine releases which were each time decreasing the evaluation.

I also write in my article that in my last match I was using the net 42820. 42850 was not available at that time yet but I expect very little or no difference at all in strength between both.

Finally don't forget that the matches were played at a rapidrate. Many well played games were spoiled by Lc0 as it was running out of time and was just living from the 10 seconds increment for many moves which clearly hurt the performance.

In both my articles I write that I would love to test Lc0 on stronger hardware but financially I can't/ won't in the nearby future. Donations are of course always welcome.
  
Back to top
WWW  
IP Logged
 
fling
God Member
*****
Offline


I Love ChessPublishing!

Posts: 1550
Joined: 01/21/11
Gender: Male
Re: New version of LcO
Reply #65 - 09/19/19 at 22:13:26
Post Tools
Furthermore, in the third game, after 17...Bxe5, 18. dxe5 is not that highly evaluated by Lc0, although the pv. But once the move is played, the evaluation at depth 27 goes to about the same advantage as SF10 at depth 35. But, yes, I agree, here Lc0 is clearly weaker than SF10.

I guess my point is that on stronger hardware and with the latest T40 net, Lc0 would likely do even better than in your match.
  
Back to top
 
IP Logged
 
fling
God Member
*****
Offline


I Love ChessPublishing!

Posts: 1550
Joined: 01/21/11
Gender: Male
Re: New version of LcO
Reply #64 - 09/19/19 at 17:35:56
Post Tools
Interesting analysis!

What net have you been using for LC0?

With Lc0 v0.22.0 and net 42850, at depth 20, the position after 11...c5 is evaluated as +0.55, and at lower depth it was never more than +0.69. After 30. Ng3, I get more around +-2.60. Do you have the evaluation set to centipawns?
  
Back to top
 
IP Logged
 
tony37
God Member
*****
Offline


I Love ChessPublishing!

Posts: 604
Joined: 10/16/10
Re: New version of LcO
Reply #63 - 09/19/19 at 14:58:51
Post Tools
brabo wrote on 09/19/19 at 06:47:46:
I wrote a second article on my blog which should normally erase any doubts why we all should use neural networks (in this case Lc0): https://chess-brabo.blogspot.com/2019/09/testing-chess-engines.html

very interesting games, thanks
  
Back to top
 
IP Logged
 
brabo
God Member
*****
Offline


Welcome chessfriend

Posts: 917
Joined: 02/02/07
Re: New version of LcO
Reply #62 - 09/19/19 at 06:47:46
Post Tools
an ordinary chessplayer wrote on 09/12/19 at 13:26:05:
brabo wrote on 09/11/19 at 17:20:56:
I finally found some time to write a small review on my blog of why we should use Lc0: https://chess-brabo.blogspot.com/2019/09/leela-lc0.html

Thanks, there's a lot of ideas in that blog post, but to my mind it does not quite explain why we should use Lc0. We can construct several related propositions:
  1. We should prepare.
  2. We should prepare with an engine.
  3. We should prepare with the latest engine.
  4. We should prepare with a specific engine.

(In the above chain, for "engine" we could substitute "book" or "database"; for the professionals additionally "second" or "team".)

For example, I think in your game against Guy Baete, you could have won by preparing the same opening without an engine, just using the database, so I don't see it as an argument in favor of Lc0 per se.

I wrote a second article on my blog which should normally erase any doubts why we all should use neural networks (in this case Lc0): https://chess-brabo.blogspot.com/2019/09/testing-chess-engines.html
  
Back to top
WWW  
IP Logged
 
IsaVulpes
Senior Member
****
Offline


No.

Posts: 289
Joined: 12/09/07
Re: New version of LcO
Reply #61 - 09/15/19 at 05:33:05
Post Tools
Keano wrote on 09/14/19 at 18:32:59:
MartinC wrote on 09/13/19 at 20:30:18:
LC0 simply won't be running very well if its on a laptop CPU.


why not?

Because the entire engine architecture is different, and the type of operations that neural nets (as opposed to classic engines like Stockfish) engage in are much more efficiently handled by a GPU than a CPU.

For Leela to show her "true" power, you need something upwards of an RTX2060 card - usable in general, especially if you let it run for a while, it is of course no matter your hardware (I have the LD2 net on my phone), but it will be significantly weaker - and most notably, much *more* significantly weaker than Stockfish becomes with dropping hardware levels.

Meaning: You can't really give an opinion on the engine's evaluations if your hardware is so poor that it barely runs. If I somehow manage to install Stockfish on my Nokia 3330, and it gets 1 node/second, going on record with "This engine is actually quite weak" becomes a bit silly.
  
Back to top
 
IP Logged
 
Keano
God Member
*****
Offline


Money doesn't talk, it
swears.

Posts: 2885
Location: Toulouse
Joined: 05/25/05
Gender: Male
Re: New version of LcO
Reply #60 - 09/14/19 at 18:32:59
Post Tools
MartinC wrote on 09/13/19 at 20:30:18:
LC0 simply won't be running very well if its on a laptop CPU.


why not?
  
Back to top
 
IP Logged
 
trw
God Member
*****
Offline


I Love ChessPublishing!

Posts: 1260
Joined: 05/06/08
Gender: Male
Re: New version of LcO
Reply #59 - 09/14/19 at 18:14:55
Post Tools
It's been the standard for over 10 years to use two engines at once... though I have never seen or used a computer strong enough to run three at any real depth.

Anyhow, disagreement usually just means that the position is sufficiently complicated and warrants more investigation... Not that one is always right.
  
Back to top
IP Logged
 
Pawnpusher
Full Member
***
Offline


I Love ChessPublishing!

Posts: 135
Joined: 01/04/18
Gender: Male
Re: New version of LcO
Reply #58 - 09/14/19 at 11:29:09
Post Tools
One of the ideas I gleaned(stole) is using three engines at the same time say Stockfish, Komodo, and LcO. If any two agree that is the best choice. If all three differ use LcO's choice. Which I guess is an argument for a desktop.
  
Back to top
 
IP Logged
 
Page Index Toggle Pages: [1] 2 3 ... 5
Topic Tools
Bookmarks: del.icio.us Digg Facebook Google Google+ Linked in reddit StumbleUpon Twitter Yahoo