Leela Chess Zero (41092) Vs Stockfish 210219 5m+5s

[mwyoung]

Lc0 hardware RTX 2080 ti
2 threads
NN=41092

Stockfish 210219 hardware
Threadripper 2950x
32 threads

TC = 5'+5''
6 move opening book
Ponder = on
6 man tablebase

Live Stream:

[mwyoung]

resultss.jpg

[Hugo]

Hi

Thanks for running this match.

a few things I noticed about this match.
First, I think your machine has 32 threads in total.
If so, Stockfish should not run on 32 threads and Leela on 2 while ponder = on.
It is because Leela in deed uses this 2 cpu threads, and to me its unclear how this affects both engines if they share threads

second, I think Leela is bit slow for running on a RTX 2080Ti card.
There are 2 parameters that might be more optimal for Leela.
--nncache=20000000
--minibatch-size=1024

kind regards, C.K.

[Chessqueen]

Lc0 hardware RTX 2080 ti
2 threads
NN=41092

Stockfish 210219 hardware
Threadripper 2950x
32 threads

TC = 5'+5''
6 move opening book
Ponder = on
6 man tablebase

Live Stream:

There is also this interesting match going on ==>
https://tcec.chessdom.com/

or on this site ===>
https://tcecbonus.club/

[mwyoung]

Hi

Thanks for running this match.

a few things I noticed about this match.
First, I think your machine has 32 threads in total.
If so, Stockfish should not run on 32 threads and Leela on 2 while ponder = on.
It is because Leela in deed uses this 2 cpu threads, and to me its unclear how this affects both engines if they share threads

second, I think Leela is bit slow for running on a RTX 2080Ti card.
There are 2 parameters that might be more optimal for Leela.
--nncache=20000000
--minibatch-size=1024

kind regards, C.K.

Thanks you for the interest. I designed this system with testing Lc0 with ponder on. And I have tested this after I built this system. The overhead of running Lc0 is very low, less then 9%. I then overclocked the system by 15%. Everything has run cool and stable running both engines with ponder on even when running the system over 24 hours in hard testing. This has worked out better then I could have hoped.

I then tested Numa vs Uma configurations and this netted the system 3% for Stockfish, and 5% for Lc0. Increase in performance.

I then tested Lc0 configurations. My configuration runs best with minibatch-size=1024. And has match benchmarks on the Lc0 site running same version and NN.

Lc0 has been testing with this configuration only, unless I find other versions and NN run best with other settings.

Unless I can find this method of testing is hurting the result and is not best. Ponder on matches is the preferred method of test always. Testing CPU only engines will be tested with ponder off. When using all threads for both engines.

I then setup the boot and configured Windows 10 to clear off all windows 10 B.S. To get the systems overhead as low as possible. My overhead at startup is 1%.

I have posted the results, testing method, and video of some of my test. Anyone if they wish can disregard the results. If they think this is hurting Stockfish, or Lc0 when testing ponder = ON.

[Hugo]

I expected you to ignore whatever comes, and its a bit frustrating to see you producing fake news.
Was the same when you gave SF 16GB hashtables for a 5 min game....geee

You invest a lot of money in your new pc, but then you ignore some simple technical fakts. Maybe ist not a huge impact in result, but check your taskmanager, then you will see that the 2 threads you give Leela are used 95% in the endgame....if not Sf wants to use them a the same time ofc.

C.K.

[mwyoung]

I expected you to ignore whatever comes, and its a bit frustrating to see you producing fake news.
Was the same when you gave SF 16GB hashtables for a 5 min game....geee

You invest a lot of money in your new pc, but then you ignore some simple technical fakts. Maybe ist not a huge impact in result, but check your taskmanager, then you will see that the 2 threads you give Leela are used 95% in the endgame....if not Sf wants to use them a the same time ofc.

C.K.

Not Ignored, but when tested and shown not to be true. It is disregarded.

If you can show an impact, please post the data.