Mystery engine at CCC

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

Moderator: Ras

Uri Blass
Posts: 10825
Joined: Thu Mar 09, 2006 12:37 am
Location: Tel-Aviv Israel

Re: Mystery engine at CCC

Post by Uri Blass »

Chessqueen wrote: Sun Aug 27, 2023 3:55 am
Graham Banks wrote: Fri Jul 14, 2023 3:24 am Announcing Torch: New #2 Chess Engine
CHESScom
CHESScom
Jul 13, 2023, 6:17 PM

Chess.com News
Torch is a brand-new chess engine built from the ground up by top chess engine developers. The team is composed of: Andrew Grant (Ethereal), Finn Eggers & Kim Kåhre (Koivisto), Jay Honnold (Berserk), and Michael Whiteley & Dietrich Kappe (Dragon). Mark Lefler and Larry Kaufman have been advisors to the project.

Many on the team have long discussed joining forces to produce an engine capable of contending with the best out there. Through the support and funding of Chess.com, we have done exactly that. A single unified effort by our developers has allowed us to rapidly produce the new number-two engine according to Torch’s performance in the Chess.com Computer Chess Championship. Our goal is to be number one.
Either Torch is getting stronger or Stockfish play some positions horrible ==>https://www.youtube.com/watch?v=3RJFt5fQZY8
I do not know the engine that lost against torch but it was certainly not stockfish or maybe some old version.

Stockfish suggest 6...Bb4 instead of 6...Qc7 and after Qc7 white is more than 0.5 pawn down.
8...Ne5 is not a move of stockfish and better is 8,,,Bd6 and later stockfish can see a clear advantage for white that is close to 1 pawn.

19.e5 and 20.Nc6 are easy to find for stockfish.
Chessqueen
Posts: 5685
Joined: Wed Sep 05, 2018 2:16 am
Location: Moving
Full name: Jorge Picado

Re: Mystery engine at CCC

Post by Chessqueen »

Uri Blass wrote: Sun Aug 27, 2023 3:18 pm
Chessqueen wrote: Sun Aug 27, 2023 3:55 am
Graham Banks wrote: Fri Jul 14, 2023 3:24 am Announcing Torch: New #2 Chess Engine
CHESScom
CHESScom
Jul 13, 2023, 6:17 PM

Chess.com News
Torch is a brand-new chess engine built from the ground up by top chess engine developers. The team is composed of: Andrew Grant (Ethereal), Finn Eggers & Kim Kåhre (Koivisto), Jay Honnold (Berserk), and Michael Whiteley & Dietrich Kappe (Dragon). Mark Lefler and Larry Kaufman have been advisors to the project.

Many on the team have long discussed joining forces to produce an engine capable of contending with the best out there. Through the support and funding of Chess.com, we have done exactly that. A single unified effort by our developers has allowed us to rapidly produce the new number-two engine according to Torch’s performance in the Chess.com Computer Chess Championship. Our goal is to be number one.
Either Torch is getting stronger or Stockfish play some positions horrible ==>https://www.youtube.com/watch?v=3RJFt5fQZY8
I do not know the engine that lost against torch but it was certainly not stockfish or maybe some old version.

Stockfish suggest 6...Bb4 instead of 6...Qc7 and after Qc7 white is more than 0.5 pawn down.
8...Ne5 is not a move of stockfish and better is 8,,,Bd6 and later stockfish can see a clear advantage for white that is close to 1 pawn.

19.e5 and 20.Nc6 are easy to find for stockfish.
I will go over it with the latest 6 differerent versions of Stockfish to determine which version was used===>https://www.youtube.com/watch?v=3RJFt5fQZY8
AndrewGrant
Posts: 1960
Joined: Tue Apr 19, 2016 6:08 am
Location: U.S.A
Full name: Andrew Grant

Re: Mystery engine at CCC

Post by AndrewGrant »

Uri Blass wrote: Sun Aug 27, 2023 3:18 pm I do not know the engine that lost against torch but it was certainly not stockfish or maybe some old version.

Stockfish suggest 6...Bb4 instead of 6...Qc7 and after Qc7 white is more than 0.5 pawn down.
8...Ne5 is not a move of stockfish and better is 8,,,Bd6 and later stockfish can see a clear advantage for white that is close to 1 pawn.

19.e5 and 20.Nc6 are easy to find for stockfish.
Yes it is Stockfish. The latest version of Stockfish that was available as of the start of the multi-day event.
Also the first 8 moves / 16 ply are opening books. So yes Stockfish would probably never play Qc7.
But both Stockfish and Torch took turns being forced to play it.

Game: https://www.chess.com/computer-chess-ch ... ls&game=24
Logs: https://cccfiles.chess.com/archive/cute ... 352135.zip
CornfedForever
Posts: 648
Joined: Mon Jun 20, 2022 4:08 am
Full name: Brian D. Smith

Re: Mystery engine at CCC

Post by CornfedForever »

302 to 198 - that's a pretty nice beat down of Team Torch
dangi12012
Posts: 1062
Joined: Tue Apr 28, 2020 10:03 pm
Full name: Daniel Infuehr

Re: Mystery engine at CCC

Post by dangi12012 »

CornfedForever wrote: Mon Aug 28, 2023 2:24 am 302 to 198 - that's a pretty nice beat down of Team Torch
Well the point would be if the S curve has an intersection in the future. Hopefully torch is written from scratch - an empty file brings fresh perspective and you are not bound to legacy approaches etc. If its a copy and paste of existing engines I dont see the intersection in the future.
Also new technologies have a higher slope in the improvement per unit time compared to the older proven technology stockfish. With the state of SF discord I would not be suprised if someone catches up sooner or later.

That being said NNUE can be made significantly faster and this is finally verified and will be made public soon on a new wiki like website. But when something is public it will improve both sides I suppose so in that regard I hope chessprogramming as a hole becomes stronger. There are many many opportunities open and worth to be explored.

For example if you really expand this into the code thats actually here when AVX2 + BMI is enabled there are many things going on - different buffers, mixing of AVX2 and SSE code and so on all with some slight tradeoffs. https://github.com/official-stockfish/S ... ure.h#L118

On that even here I can repeat and say its worth to serialize weights and biases into a single uniform aligned buffer so that each time any weight or bias is accessed inside nnue eval it becomes the next element in already preloaded array.
Changing the dependency on memory and number of unique pointers for the compiler to consider inside inference after the accumulator from
F(fc_0_w, fc_0_b, fc_1_w, fc_1_b, fc_2_w, fc_2_b)
to
F(wb)
where we can just *wb++ any time we access weights or biases.
Of course that needs a complex initialisation machinery (which already is in there) since depending on simd architecture you need to shuffle and align weights too as well as correct padding etc. But that alone makes the precious runtime inference a tick faster
Worlds-fastest-Bitboard-Chess-Movegenerator
Daniel Inführ - Software Developer
Jouni
Posts: 3621
Joined: Wed Mar 08, 2006 8:15 pm
Full name: Jouni Uski

Re: Mystery engine at CCC

Post by Jouni »

I don't believe Torch can beat SF. How can You beat team having tens of brilliant programmer with 6 men. No way!
Jouni
smatovic
Posts: 3296
Joined: Wed Mar 10, 2010 10:18 pm
Location: Hamburg, Germany
Full name: Srdja Matovic

Re: Mystery engine at CCC

Post by smatovic »

Jouni wrote: Sat Sep 02, 2023 2:04 pm I don't believe Torch can beat SF. How can You beat team having tens of brilliant programmer with 6 men. No way!
Tens of brilliant - voluntary - programmers vs. 6 men who - might - get some compensation?

***edit***

Team OpenBench + team Komodo vs. team SF + team Lc0 vs. team Rebel + CSTal, fun!

--
Srdja
Chessqueen
Posts: 5685
Joined: Wed Sep 05, 2018 2:16 am
Location: Moving
Full name: Jorge Picado

Re: Mystery engine at CCC

Post by Chessqueen »

Jouni wrote: Sat Sep 02, 2023 2:04 pm I don't believe Torch can beat SF. How can You beat team having tens of brilliant programmer with 6 men. No way!
Believe it, here is another win by Torch https://www.youtube.com/watch?v=vuq3y3QRkVk
DmitriyFrosty
Posts: 58
Joined: Mon Mar 27, 2023 8:29 pm
Full name: Dmitry Frosty

Re: Mystery engine at CCC

Post by DmitriyFrosty »

Torch managed to draw against SF in CCC 21 Rapid Quaterfinals (12 engines playing each other). Very nice result by Torch team, hope you will dethrone SF 👍.