Big-SF 200324

Discussion of anything and everything relating to chess playing software and machines.

Moderator: Ras

Eduard
Posts: 1439
Joined: Sat Oct 27, 2018 12:58 am
Location: Germany
Full name: N.N.

Big-SF 200324

Post by Eduard »

After getting a lot of letters from all over the world and some other people also being dissatisfied with Stockfish dev, I decided to release another engine. The meaning is: The influence of the small network was reduced and the large network was implemented with dimension 3072 (instead of dimension 2560). I use this engine for Blitz and Bullet as well as for quick analysis. In my opinion it is clearly better than SF dev for short analysis, and maybe (just maybe?) -2 Elo in bullet.

(My critics here in the forum can now come and scoff. Have fun with it, but I'm telling you, it's not me, it's you who are making a fool of yourself!)

By the way, search.cpp has been slightly modified and implemented "Minimum Thinking Time" and "Slow Mover". I implemented MTT again because I recently lost a game where the engine moved in 0s. This didn't happen with less than 1s on the clock but with more time. There are also positions where you need to think longer. That's why I'm going to implement MTT again. I've been playing like this for several years and haven't losed a game because of moves in 0s.

Big-SF 200324 Win 64-Bit (avx2, avx512, bmi2, sse41, source code, external networks):
https://pixeldrain.com/u/xK1DY2NH

Have fun!
Eduard
Posts: 1439
Joined: Sat Oct 27, 2018 12:58 am
Location: Germany
Full name: N.N.

Re: Big-SF 200324

Post by Eduard »

Here you can also download and install other networks with Dimension 3072 using the engine. A new D 3072 net is currently being tested almost every day: https://tests.stockfishchess.org/nns

All you have to do is look at the following line:
like here: https://tests.stockfishchess.org/tests/ ... 0526c4fbd9

// Number of input feature dimensions after conversion
-constexpr IndexType TransformedFeatureDimensionsBig = 2560;
+constexpr IndexType TransformedFeatureDimensionsBig = 3072;
constexpr int L2Big = 15;
constexpr int L3Big = 32;
Draude
Posts: 31
Joined: Sun Mar 10, 2024 1:56 pm
Full name: Draude Stoalis

Re: Big-SF 200324

Post by Draude »

Hello Eduard! THANK YOU so much for releasing this Big-SF version! Dimension 3072 has needed to be added for a long time now... But I guess Stockfish devs never listen to logic until I show them the results of a 1 billion games test! I will be switching my main analysis engine now. Soon, I will also post my results against the previous Big-SF version.

Are there any plans to add your own dimension 4096 nets to the engine?

By the way, are CorChess and Leptir-Big-Ultra still being developed by you? Can I have the binaries to run tests with my test suites? Thanks.
Pali
Posts: 30
Joined: Wed Dec 01, 2021 12:23 pm
Full name: Doruk Sekercioglu

Re: Big-SF 200324

Post by Pali »

Eduard wrote: Wed Mar 20, 2024 8:34 am After getting a lot of letters from all over the world and some other people also being dissatisfied with Stockfish dev, I decided to release another engine. The meaning is: The influence of the small network was reduced and the large network was implemented with dimension 3072 (instead of dimension 2560). I use this engine for Blitz and Bullet as well as for quick analysis. In my opinion it is clearly better than SF dev for short analysis, and maybe (just maybe?) -2 Elo in bullet.

(My critics here in the forum can now come and scoff. Have fun with it, but I'm telling you, it's not me, it's you who are making a fool of yourself!)

By the way, search.cpp has been slightly modified and implemented "Minimum Thinking Time" and "Slow Mover". I implemented MTT again because I recently lost a game where the engine moved in 0s. This didn't happen with less than 1s on the clock but with more time. There are also positions where you need to think longer. That's why I'm going to implement MTT again. I've been playing like this for several years and haven't losed a game because of moves in 0s.

Big-SF 200324 Win 64-Bit (avx2, avx512, bmi2, sse41, source code, external networks):
https://pixeldrain.com/u/xK1DY2NH

Have fun!
Could you point to a source for "... it is clearly better than SF dev for short analysis ..." and the "... and maybe (just maybe?) -2 Elo in bullet."?
bastiball
Posts: 5356
Joined: Tue Oct 20, 2020 4:18 am
Full name: Basti Dangca

Re: Big-SF 200324

Post by bastiball »

I'll test it in FRC if there's a difference :D
Basti Dangca
CCRL testing group
Damir
Posts: 2859
Joined: Mon Feb 11, 2008 3:53 pm
Location: Denmark
Full name: Damir Desevac

Re: Big-SF 200324

Post by Damir »

You better remember to set a path to engine nnue. The engine is not embeeded.. If you don't set a path to nnue, the engine will show an error message '' Engine Caused Exception'' and game will end after few moves. You must set full path to nnue, where it is located, to avoid any errors.
Eduard
Posts: 1439
Joined: Sat Oct 27, 2018 12:58 am
Location: Germany
Full name: N.N.

Re: Big-SF 200324

Post by Eduard »

22. Mar 2024: Big-SF 210324. Latest SF updates and improved search for the large network. As I already wrote, the SF developers are testing new D 3072 networks. Here is a test version with the network "nn-56e7225a7727.nnue". The network is now embedded because some users are having problems with external networks (even though it's easy). Features: Slow Mover, Minimum Thinking Time. Download Win 64-Bit (avx2, bmi2, sse41, source code): https://pixeldrain.com/u/vqWuhCtX
Eduard
Posts: 1439
Joined: Sat Oct 27, 2018 12:58 am
Location: Germany
Full name: N.N.

Re: Big-SF 200324

Post by Eduard »

Big-SF 220324 (dimension 3072). There is a new developer test version with new network "nn-31636a26c4f6.nnue". I haven't changed anything this time, but I have implemented Minimum Thinking Time and Slow Mover for everyone who wants to play online.

Changed developer code:
https://github.com/official-stockfish/S ... f7c29444d4

The version with Dimension 3072 network against D 2560 is currently being tested on Fishtest, live:
https://tests.stockfishchess.org/tests/ ... 0526c52815

Download 64-bit (avx2, bmi2, sse41, source code):
https://pixeldrain.com/u/NMM14eoP

Have fun testing.
bastiball
Posts: 5356
Joined: Tue Oct 20, 2020 4:18 am
Full name: Basti Dangca

Re: Big-SF 200324

Post by bastiball »

Not bad in FRC

Code: Select all

Score of Big-SF 200324 vs Stockfish 16.1: 7 - 16 - 77 [0.455]
...      Big-SF 200324 playing White: 5 - 7 - 38  [0.480] 50
...      Big-SF 200324 playing Black: 2 - 9 - 39  [0.430] 50
...      White vs Black: 14 - 9 - 77  [0.525] 100
Elo difference: -31.4 +/- 32.4, LOS: 3.0 %, DrawRatio: 77.0 %
SPRT: llr 0 (0.0%), lbound -inf, ubound inf
Basti Dangca
CCRL testing group