Stockfish NNue [download ]

[Rebel]

the latest well tested version is now ~95 elo stronger than Stockfish 11

Worth noting: this is at 40 moves in 20 seconds time control. At real world time controls like 5 +0 I don't think 70 elo has been broken yet (but... this is more than one year of progress still...)

And you likely missed that a couple of days ago I played 2000 games at CCRL Blitz (that's a lot more games than CCRL) and that the result was even better than the 40m/20s run.

[Ovyron]

And you likely missed that a couple of days ago I played 2000 games at CCRL Blitz (that's a lot more games than CCRL) and that the result was even better than the 40m/20s run.

Adam Hair has shown that such "n moves in n time" might be unreliable because the time used per move draws some weird curves and the engine is forced to adjust its time and use time where it doesn't matter, or not have enough time where it matters.

Here's a graph:



The engine gets a lot more time at move 40 for no reason at all. I suggest you retest at some fixed time per game or fixed time with increment time controls.

[Damir]

https://abrok.eu/stockfish-nnue/

[cma6]

Rebel, thank you for providing a great service to the chess engine community. To keep updated on the best SF NNUE version, is it sufficient to update to the latest (non-experimental network (.bin) from Sergio, or should one also update to your latest version of "sf-nnue-bmi2.exe"?

The latest Sergio net may not be the strongest one, let others test and upgrade once someone you trust reports that there's a new net that is the best.

Thanks, Ovyron, for the advice. May I ask who you follow for testing and reporting on the best of Sergio's nets?

[MMarco]

How do can you test the speed of the compile to see which compile is fastest on your system? Is there a console command?

Try the bench command: https://github.com/mcostalba/Stockfish/ ... chmark.cpp

/// bench -> search default positions up to depth 13
/// bench 64 1 15 -> search default positions up to depth 15 (TT = 64MB)
/// bench 64 4 5000 current movetime -> search current position with 4 threads for 5 sec
/// bench 64 1 100000 default nodes -> search default positions for 100K nodes each
/// bench 16 1 5 default perft -> run a perft 5 on default positions

[Paul Bedrey]

I'm curious what the graph on Sergio's download page means. It's a rating of sorts.

Seems like there is a need of a thread devoted for posting user results.

I am running games by hand on two separate machines. Based on ponder move my impression is that NNue acts as a totally different engine from StockFish. I'm wondering what a similarity test would show.

[Jouni]

I am curious: how does NNue score in Your speedy rating list and 40 000 positions?!

[h1a8]

the latest well tested version is now ~95 elo stronger than Stockfish 11

Worth noting: this is at 40 moves in 20 seconds time control. At real world time controls like 5 +0 I don't think 70 elo has been broken yet (but... this is more than one year of progress still...)

And you likely missed that a couple of days ago I played 2000 games at CCRL Blitz (that's a lot more games than CCRL) and that the result was even better than the 40m/20s run.

Question
So will the engine be NEARLY the same strength regardless of the graphics card?

For example, I have a 2080ti with i7-9700.
Will the engine be NEARLY the same strength if I just had a 1650ti
with i7-9700?

[Ovyron]

Question
So will the engine be NEARLY the same strength regardless of the graphics card?

For example, I have a 2080ti with i7-9700.
Will the engine be NEARLY the same strength if I just had a 1650ti
with i7-9700?

The engine doesn't use a graphics card at all, so it's not a concern.

[Ovyron]

Thanks, Ovyron, for the advice. May I ask who you follow for testing and reporting on the best of Sergio's nets?

I've been following the tests of people from this thread, though realistically you'd only want to download a net when it's clearly stronger than one you already have.