SPCC: Testrun of Slow Chess 2.8 finished

[pohl4711]

Ratinglist-testrun of Slow Chess 2.8 finished.

https://www.sp-cc.de

(Perhaps you have to clear your browsercache or reload the website)

[mehmet123]

Great performance for Slow Chess 2.8. It's now ahead Koivisto 7.5 avx2 and Ethereal 13.25 NNUE.
Slow Chess ranks 3rd among alpha beta engines using its original network and it uses only a 3.5 Mb size NNUE.

[pohl4711]

Great performance for Slow Chess 2.8. It's now ahead Koivisto 7.5 avx2 and Ethereal 13.25 NNUE.
Slow Chess ranks 3rd among alpha beta engines using its original network and it uses only a 3.5 Mb size NNUE.

Yes, the progress and strength-level of Slow Chess 2.8 is impressive.
What IMHO is very interesting about Slow Chess is, that Slow Chess uses more than one neural-net. It has one net for the middlegame and several nets for endgame as far as I understand the Changelog...
Slow Chess 2.8 Changelog:
More neural net training.
Updated endgame nets. (Training, more first layer weights, less second, more forced symmetry.)
Larger general net.

[jonkr]

Thanks for testing, Slow always seems to do well in SPCC, so I'm happy to see it has exceeded my expectations. I think Slow likes the openings, and the large number of games played makes results more consistent.

For general evaluation it uses a 2-part net, the main part is all the piece inputs updated incrementally, and the second part is Pawn-King-Material-CastleRights net which caches its outputs in a hash table that is combined in with the general net. The nets are all 16-bit values so if I ever try out further quantizing values data size might get smaller even.

All the other nets are smaller endgame nets, mostly for specific endgames like R v R, Q v Q, B v B, N v N, B v N, R v N, R v B, Q v Rs (with any number of pawns on the board.)

[Jouni]

Really stunning engine. Beats Koivisto 7.5 clearly in my test. And this small engine has also built in GUI .