I wanted to test how good Java is at evaluating neural networks, so I ported the NN code I use in Texel to CuckooChess. As expected the performance is really bad compared to hand-optimized C++ code using SIMD intrinsics.
Consequently CuckooChess 1.13 is 10 to 15 times slower (NPS) than the previous CuckooChess version, and 20 to 25 times slower than Texel. It is still quite a bit stronger than the previous CuckooChess version, at least at bullet and blitz time controls.
Maybe some people will find it interesting to have an engine with a decent evaluation but very weak tactical abilities, so I am releasing this test as CuckooChess 1.13. Here is the download link:
CuckooChess 1.13
CuckooChess 1.13
Moderators: hgm, Rebel, chrisw
-
- Posts: 697
- Joined: Mon Apr 19, 2010 7:07 pm
- Location: Sweden
- Full name: Peter Osterlund
-
- Posts: 82
- Joined: Sat Aug 01, 2015 6:16 pm
- Location: France
- Full name: Eric Bonneau
Re: CuckooChess 1.13
Hi Peter,
nice, thanks for sharing it!
BR,
Eric
nice, thanks for sharing it!
BR,
Eric
-
- Posts: 58
- Joined: Mon Mar 27, 2023 8:29 pm
- Full name: Dmitry Frosty
Re: CuckooChess 1.13
Nice gift. Waiting for Texel update as well. Thank you.
-
- Posts: 4495
- Joined: Fri Apr 21, 2006 4:19 pm
- Location: IASI - the historical capital of MOLDOVA
- Full name: SilvianR
Re: CuckooChess 1.13
Thank you very much for the gift ! I have already added it to my (ongoing) Java ELITE Tournament (TC=4'+2"). CuckooChees 1.12 x64 is also playing in this tournament. Serendipity-0.1 (the current authoritative leader) is also a NNUE engine. And Bagatur 4.1a has a NNUE version (but the neural network belongs to Stockfish, so I use the HCE version in the tournament).Details here:petero2 wrote: ↑Wed May 01, 2024 11:12 am I wanted to test how good Java is at evaluating neural networks, so I ported the NN code I use in Texel to CuckooChess. As expected the performance is really bad compared to hand-optimized C++ code using SIMD intrinsics.
Consequently CuckooChess 1.13 is 10 to 15 times slower (NPS) than the previous CuckooChess version, and 20 to 25 times slower than Texel. It is still quite a bit stronger than the previous CuckooChess version, at least at bullet and blitz time controls.
Maybe some people will find it interesting to have an engine with a decent evaluation but very weak tactical abilities, so I am releasing this test as CuckooChess 1.13. Here is the download link:
CuckooChess 1.13
==================
viewtopic.php?p=963543#p963543
==================
-
- Posts: 697
- Joined: Mon Apr 19, 2010 7:07 pm
- Location: Sweden
- Full name: Peter Osterlund
Re: CuckooChess 1.13
Thanks for testing. I actually had your tournament in mind when deciding to release this version.Sylwy wrote: ↑Wed May 01, 2024 4:49 pmThank you very much for the gift ! I have already added it to my (ongoing) Java ELITE Tournament (TC=4'+2"). CuckooChees 1.12 x64 is also playing in this tournament. Serendipity-0.1 (the current authoritative leader) is also a NNUE engine. And Bagatur 4.1a has a NNUE version (but the neural network belongs to Stockfish, so I use the HCE version in the tournament).Details here:petero2 wrote: ↑Wed May 01, 2024 11:12 am I wanted to test how good Java is at evaluating neural networks, so I ported the NN code I use in Texel to CuckooChess. As expected the performance is really bad compared to hand-optimized C++ code using SIMD intrinsics.
Consequently CuckooChess 1.13 is 10 to 15 times slower (NPS) than the previous CuckooChess version, and 20 to 25 times slower than Texel. It is still quite a bit stronger than the previous CuckooChess version, at least at bullet and blitz time controls.
Maybe some people will find it interesting to have an engine with a decent evaluation but very weak tactical abilities, so I am releasing this test as CuckooChess 1.13. Here is the download link:
CuckooChess 1.13
==================
viewtopic.php?p=963543#p963543
==================
-
- Posts: 4495
- Joined: Fri Apr 21, 2006 4:19 pm
- Location: IASI - the historical capital of MOLDOVA
- Full name: SilvianR
Re: CuckooChess 1.13
-
- Posts: 1247
- Joined: Sat Jul 05, 2014 7:54 am
- Location: Southwest USA
Re: CuckooChess 1.13
One Of My Personal Favorites (A Great Name)......Not easy to get to work...but well worth the time!! Thanks!!petero2 wrote: ↑Wed May 01, 2024 11:12 am I wanted to test how good Java is at evaluating neural networks, so I ported the NN code I use in Texel to CuckooChess. As expected the performance is really bad compared to hand-optimized C++ code using SIMD intrinsics.
Consequently CuckooChess 1.13 is 10 to 15 times slower (NPS) than the previous CuckooChess version, and 20 to 25 times slower than Texel. It is still quite a bit stronger than the previous CuckooChess version, at least at bullet and blitz time controls.
Maybe some people will find it interesting to have an engine with a decent evaluation but very weak tactical abilities, so I am releasing this test as CuckooChess 1.13. Here is the download link:
CuckooChess 1.13
-
- Posts: 82
- Joined: Sat Aug 01, 2015 6:16 pm
- Location: France
- Full name: Eric Bonneau
Re: CuckooChess 1.13
Hi Peter,
just pointing out the 1.13 version tag was already used, somehow I have it on Android (1.13a9 to be fully accurate, seems to be from 2011... 13 years ago).
By the way, was CuckooChess 1.13 NN built from self-play, or is it Texel one? Or sourced elseway?
Thanks again,
Warm regards,
Eric
-
- Posts: 697
- Joined: Mon Apr 19, 2010 7:07 pm
- Location: Sweden
- Full name: Peter Osterlund
Re: CuckooChess 1.13
That is because of the non-standard version scheme I use. 1.13a9 means 1.13 alpha 9, which means a non feature complete version that will after more improvements become the real 1.13 version.Tibono wrote: ↑Tue May 07, 2024 8:03 amHi Peter,
just pointing out the 1.13 version tag was already used, somehow I have it on Android (1.13a9 to be fully accurate, seems to be from 2011... 13 years ago).
By the way, was CuckooChess 1.13 NN built from self-play, or is it Texel one? Or sourced elseway?
Thanks again,
Warm regards,
Eric
The network is exactly the same as in Texel 1.12a4. All training data for the Texel network comes from self-play games. The trainer is 100% written by me, and the evaluation code is also 100% written by me.
There is no github repository for CuckooChess 1.13 because CuckooChess is part of the DroidFish repository, which is already larger than desired because of the huge Stockfish nets. The source code for CuckooChess 1.13 is however included in the jar file in case someone wants to look at it.