Interessanter Blog-Artikel

https://blog.lczero.org/2018/07/04/leela-ratio/

Hier der Wortlaut:

"Leela Ratio: What is it?

If you spend any amount of time around the Leela forums and chats, you’ll see an announcement about a gauntlet or match where IDXXX has defeated some vaunted engine. Maybe there’ll be a note about the time control and the GPU or the number of cores the opponent used. But you are often left scratching your head as to how strong or weak Leela was relative to their opponent.

Enter the Leela Ratio: LeR = 875 * leela nps/sf9 nps

This is supposed to be one number that encapsulates how powerful the CPU is versus the GPU. Its derived from the reported ratio between Alpha Zero’s and Stockfish 8’s nodes per second – 80k and 70m – during their match. So Stockfish 8 was searching 875 times as many nps as A0. Plugging those figures into the formula you get 1.0. Anything higher than that, and the GPU is stronger, less than 1.0, and the CPU is stronger.

How does one get the nps for Leela and Stockfish 9 (most people have 9 instead of 8)?
Leela NPS

Quoting from the lc0 benchmark wiki page.

“Run go infinite from start position and abort after depth 26 and report NPS output.”
SF9 NPS

Fire up the sf9 engine, “go infinite” from the start position, and around depth 26-28, abort and note the nps.

As an example, on my laptop with a GTX 1070 and 1 core for sf9, I get around 10,000 nps with a 192x15 net and 1,410,000 nps for sf9. Plugging that into our equation, we get LeR = 6.2, so a bit in favor of the GPU."

Ich habe den Leela-Ratio-Wert natürlich gleich mal für mein long thinking-time Turnier, wo ich ja Leela (Cuda-Version) teste(https://www.sp-cc.de/long-thinkingtime.htm), ausgerechnet und erfreut festgestellt, daß meine relativ langsame Grafikkarte (GTX 950m) in Kombination mit den im singlethread-Modus spielenden Gegner-Engines, auf meiner Hardware einen Leela-Ratio Wert von 1.5 ergibt, was sehr nahe an 1.0 liegt (was der Wert beim AlphaZero-Match gegen Stockfish 8 war (s.o.)). Paßt! Würde allen, die Leela testen, empfehlen, den Ratio-Wert ebenfalls mal auszurechnen und, falls er stark von 1 abweicht (2 oder 3, oder gar noch höher), die jeweiligen Testbedingungen entsprechend anzupassen (also den Gegner-Engines mehr oder weniger CPU-Cores/Threads zuzuweisen).