TalkChess.com

Some Ceres comments: Yes, I did get it working. No, I have no idea what combination of incantations made it work. Yes, doing so is a major PITA. Clearly, Ceres was a huge amount of work. The good news: the developer seemed intent on a New Year's release on the Leela Chess Zero Discord server. The ba...

Looks like things are a bit premature for using Ceres. Different "node" definitions render initial fixed node tests vs Lc0 invalid with default parameters, and time management problems result in time forfeits so my testing is on hold until the issues are addressed. See discussions on Leela Chess Zer...

You are correct.
My mistake.

Grant's results look like perfectly fine scaling to me...
What am I missing?

Allow me to disagree.
Since the games are the same engine against itself, your matches do not show anything directly about scalability, other than perhaps a second order effect. The lower margins of superiority might simply be a result of more draws.

Following the link to your tests I noticed you were playing the same engine against itself I think.
My match was SF-NNUE vs SF11, so they are not the same.
Might that account for the different results?

Ah, OK.
I'll look forward to seeing the table of your matches 1, 2, 4, and 8 CPUs then.
A different result would be interesting.

The engine used was SF with NNUE.
It scales fine.
It does not matter how much it won by, just that it scaled reasonably.
I don't know about other versions.

Engine was stockfish.avx2.halfkp_256x2-32-32.profile-nnue.2020-07-19.exe
Net was gekkehenker2020-06-27.bin

This was the match command used for the 1 cpu games (IIRC): cutechess-cli -each tc=0:20+1.0 restart=on -engine conf=SF11 -engine conf=SF-NNUE -games 12000 -wait 1000 -ratinginterval 10 -tb E:/syzygy -tbpieces 6 -openings file=D:/Cutechess-cli/book_3moves_cp1-24_10944pos.pgn start=1 -repeat -concurre...