Meet Zahak 7 (NNUE)
I am happy to announce Zahak 7.0. It it is the first version of Zahak with NNUE, yup you read it right, NNUE. With the help of Aryan Parekh (the author of Bit-Genie), I wrote an NNUE trainer and produced a dataset of around 58 million chess positions using a tool that I wrote. All the chess positions were from self-play games of Zahak 6.2.
The cool thing about the network is that, in theory they will work just fine on the upcoming Bit-Genie versions, as we agreed to support the same architecture and binary format, in an effort to create some standardization.
For 64-bit architectures, I only provide AVX binaries (as the none-AVX ones were severely slower), but I also provide 32-bit and ARM bulds as usual. Note that I have spent no effort on making the ARM and 32-bit builds fast, as who uses those architectures often don’t care much about the sheer strength, but rather an engine to play with. That also includes M1 devices, which are not meant for performance but power efficiency.
Zahak 7 comes with +190 elo points over Zahak 6.2, and from my tests with other engines (CounterGo 3.7 and Bit-Genie 9) I estimate it to be around 2960 Elo points (as per CCRL).
You can actually help and train nets for Zahak, here is a guide for you, what is cool if your net is chosen to be promoted to a master net, you will be listed as an author of Zahak!
Apart from NNUE, a few more changes are also included:
- Fix reporting mate scores
- Removing futility pruning in quiescence search
- Various search parameter tuning to adapt it to NNUE
- Exposing EvalFile and BookPath UCI options
- NPS improvement
Still to come:
- Include different networks, for different skill levels.
- Support endgame table bases.
- Support for searchmoves and go mate commands.
To download Zahak 7.0, please visit here
Enjoy playing with Zahak!