LC0-784803-0.29.0.Rc vs Sf15 (50sec+1sec)

[pawel....@gmail.com]

I havent been running this kind of tournament for long time, but today, just for the sake of curiosity I decided to test the newest 784803 net 3080 Rtx vs SF15 on 10 cores. Book on up to 8 moves, tested on Fritz 17 GUI. To my surprise, LC0 beat SF by small margin on 1 game 65.5-64.5 130 games played. That is really something because earlier, couple of months ago I tested some nets versus weaker Sf 12, 13, 14 and they were unable to get even close to Sf running even on 6 or 4 cores . So now I see that the progress has been made, Big progress. Good job Leela guys. I think this new versions may finally claim the throne in TCEc leagues . I hope soon they will....

[pawel....@gmail.com]

[DBg]

I am a total ignorant (with some bliss) about engine tournament. But I have been under the impressions that the categories of competition has not changed since before A0 appeared on the horizon, and to this date.  Could someone contradict that last statement for me, by mentioning, categories where NN inherent parallelism is fully exploited to reduce the full position information evaluation CPU computing cost?

I look at CCRL site. and i only see CPU and core categories. that might be nice for big threads (i guess) but not a the fine grain parallelism of NN which could literally in the extreme could be a one step paralel input-output fonction of a position. for the few position of type B-ish engine that full NN engine like lc0 typically need to explore to make its decision.

Somehow the quality of the full static evaluation, is not being put into the competition.  It still seems to be about fast tree search, not complete tree search, or accurate evaluation. (forget the chess accuracy ill-defined concept, i talk about accuracy of evaluation as estimator of true odds for example, bias versus random error for example).

Please let me me know that I have it all wrong.. that tournaments have evolved. that there are different characteristics of engine if they could be separated into tree search and evaluation modules for example, and be made to compete on each independently, as opposed to the blend we have now (in my probably obsolete understanding).

TCEC does have some GPU category, I thought. but does it has many GPUs.. what is the limite.. What are the tournament that are the most parallelism aware and accepting?

is this the right place to ask? I thought so, because in the op I only saw 10 , 5, 3, corre. not the number of GPUs  but for one 3080 Rtx presence.. I may also not be realistic about parallelism with GPUs.  but feel free to adjust.