SPCC: Testrun of Lc0 384x30-t40-1705 finished

[pohl4711]

NN-longtime- testrun of Lc0 0.23.1 384x30-t40-1705 (S.Vieri) finished (300 SALC-Armageddon games with 8'+5'' vs Stockfish 191210, average game-duration 30 minutes! Testrun-duration: 6 days).

Important news: The bullet-NN-testing is discontinued... Check out the excellent new bullet-ratinglist of Andreas Strangmueller instead:
http://www.fastgm.de/16-60-0.6.html
I decided to discontinue my bullet-NN-testing, because Andreas has a faster hardware and does nearly the same testing, as I did. So, for me, it makes no sense, to continue my testings. Because my hardware is slower (notebooks), it makes much more sense to continue my NN-testings with longer thinking-time. Checkout the new "NN longtime testing"- section of my website...
The "Lc0 / NN testing"- section (bullet NN-testings) will stay online for 1-2 months and then, I will delete it... So, if you want to download the played games, do it right now!


https://www.sp-cc.de

(Perhaps you have to clear your browsercache or reload the website)

[Hai]

https://www.sp-cc.de/nn-longtime-testing.htm
GPU: 1x mobile RTX 2060 (It is based on the desktop RTX 2060 chip but at reduced GPU clock rates (-30%, or even more when heat kicks in))
Time control: 8' + 5''

1 Lc0 0.23.1 LS 12.2 (20x256) : 3562 300 (+153,= 0,-147), 51.0 %
2 Lc0 0.23.1 t40-1705 (30x384) : 3548 300 (+147,= 0,-153), 49.0 %
3 Lc0 0.23.1 42741 (20x256) : 3541 300 (+144,= 0,-156), 48.0 %
4 Lc0 0.23.1 J13B.2-200 (24x320) : 3529 300 (+139,= 0,-161), 46.3 %

Time control per move was to short and GPU is to weak for LC0 Sergio 30x384 nets.
But when using at least 1x desktop RTX 2060, LC0 Sergio 30x384 should be on first place:)
That's really amazing that only 1 GPU and also the weakes desktop GPU RTX 2060 is enough for the LC0 Sergio 30x384 net.
= 1x desktop RTX 2080 Ti or 1x RTX Titan is good enough for the 40x512 LC0 Sergio nets.
We looking also forward to see Leelenstein 40x512 nets.

[jjoshua2]

Edosani is streaming with 2 RTX GPUs and the 24b and 30b nets are struggling. It seems that 20b are at least as good and more likely better in his blitz testing already a few minutes longer than bullet.

[mwyoung]

Edosani is streaming with 2 RTX GPUs and the 24b and 30b nets are struggling. It seems that 20b are at least as good and more likely better in his blitz testing already a few minutes longer than bullet.

If they are struggling. I wonder if they are being run on default setting. Default is not the best for the big nets.

[jjoshua2]

I think he is using defaults for all binaries. Someone should really make some PRs for better defaults I guess.
Has anyone done a test of new defaults that is good enough to get such a PR merged?

[mwyoung]

I think he is using defaults for all binaries. Someone should really make some PRs for better defaults I guess.
Has anyone done a test of new defaults that is good enough to get such a PR merged?

That explains why he is struggling. I have posted here that you must change from default when running the big nets.

I test with better settings then default. And you can see my results here.... I don't hide my settings.