Stockfish 210720 with NNUE

[kranium]

For convenience, a recent Stockfish dev with NNUE support
and a UCI option to turn NNUE eval 'on' or 'off'

https://github.com/FireFather/sf-nnue/releases

[Werner]

Thanks,
which nets are supported?

[kranium]

Thanks,
which nets are supported?

It's halfkp-256

[Werner]

Ok, yes just tested,
384x2 net is not supported - will you think about changing that?;
...and is the build fast enough compared to nodchip compiles?

[kranium]

Ok, yes just tested,
384x2 net is not supported - will you think about changing that?;
...and is the build fast enough compared to nodchip compiles?

Hi Werner-
I don't really think there's a need for 384 at this point...almost all the 256 nets I seen are more than adequate.

Whether they're fast enough is something I can't answer, a value judgement. I'll let you decide that for yourself.

My intention here was to simply figure out how to include both evals in one binary.
I would normally submit the idea as a PR to github repository, but it's been 'frozen' for the time being.

Best regards,
Norm

[perejaslav]

Not working in Fritz GUI (

[kranium]

Not working in Fritz GUI (

Hmmm too long (character limit) id name was an issue with fritz/chessbase...

I'll bet it's the same problem with id author :
T. Romstad, M. Costalba, J. Kiiski, G. Linscott, H. Noda, Y. Nasu, M. Isozaki
yes 3 have been added

I'll compile some new ones...but not right away, maybe tomorrow

Thanks for letting me know!

[Rebel]

Also:

option name EvalNNUE type check default false

Set the default setting to true.

[kranium]

shorter id author field (for compatibility with Fritz/Chessbase)
UCI option EvalNNUE default = true

https://github.com/FireFather/sf-nnue/r ... g/07-22-20

Note: I have not tested with Fritz (don't own it).

[kranium]

sf-nnue aio 07-22-2020 available
https://github.com/FireFather/sf-nnue/releases

stockfish+nnue aio (all-in-one)

All nnue modules (GEN, NNUE, LEARN) compiled into a single executable
w/ UCI options for EvalNNUE (if 'off' regular Stockfish evaluation is used) and UseEvalHash

The download includes optimized bmi2, avx2, & popc binaries