87% of the poll voted that SF will win Tcec Sufi 19.

[Nay Lin Tun]

Some people even say Leela winning Sufi is impossible.


And now it is still tied for 26 games so far.
https://cdn.discordapp.com/attachments/ ... sufi19.png


Reasons?

[Damir]

Who cares about Leela. Lc0 lost previous EMT from Stockfish by 9-10 games difference.. I would not talk much if I were you.....

[Thomas Lagershausen]

The times they are a changin.

[mwyoung]

Some people even say Leela winning Sufi is impossible.


And now it is still tied for 26 games so far.
https://cdn.discordapp.com/attachments/ ... sufi19.png


Reasons?

Yes, it is possible. But the edge must be given to Stockfish. Due to the opening bias present with TCEC SUFI. Stockfish is a better pure tactician. And a +.50cp to over a +1cp bias in TCEC openings will clearly favor Stockfish.

[RogerC]

It's like comparing a car how needs a V12 8L engine vs a 2L 4 cylinder car, just to have same speed. Stockfish is just ultra poweful with only 1 CPU. It can beat LC0 who needs more than 2000 Cuba cores graphics card to play correctly. I used to buy big graphics card just to have LC0 working, but now I just have a laptop computer with 8 CPU core, very small graphics card, and SF is the best !

LC0 idea is finished, just an old fashion NN chess engine that did not evolved. The best now is A/B search engines with standard and NNUE evaluations mixed together and tuned to find the best evaluation for a given position.

[mwyoung]

It's like comparing a car how needs a V12 8L engine vs a 2L 4 cylinder car, just to have same speed. Stockfish is just ultra poweful with only 1 CPU. It can beat LC0 who needs more than 2000 Cuba cores graphics card to play correctly. I used to buy big graphics card just to have LC0 working, but now I just have a laptop computer with 8 CPU core, very small graphics card, and SF is the best !

LC0 idea is finished, just an old fashion NN chess engine that did not evolved. The best now is A/B search engines with standard and NNUE evaluations mixed together and tuned to find the best evaluation for a given position.

You could be right. But SF is a hybrid engine for a reason. If Stockfish can increase its NN size and power with only CPU power. Then yes, Lc0 will have a hard time.

But again who is to say. That Lc0 or others engines will not use both CPU, and GPU in the future. Horsepower is Horsepower. If you get my meaning.

[yurikvelo]

LC0 idea is finished, just an old fashion NN chess engine that did not evolved.

both NN and A/B have big potential for development

Currently, A/B development speed is better.
limited resources (both for Fishtest and NN generation vs LC0 training networks) currently favour Fishtest.
Not many contributors to LC0 (quite hard to donate, you need powerful and power-hungry GPU and willing to donate it completely) and many competitive network runs (different projects, different network sizes). To develop TCEC-grade top networks (30x384) even less contributors.

https://training.lczero.org/active_users
https://tests.stockfishchess.org/users/monthly

[RogerC]

You could be right. But SF is a hybrid engine for a reason. If Stockfish can increase its NN size and power with only CPU power. Then yes, Lc0 will have a hard time.

But again who is to say. That Lc0 or others engines will not use both CPU, and GPU in the future. Horsepower is Horsepower. If you get my meaning.

Yes you're right : futur could be A/B search + standard evaluations on CPU + and bigger NNUE which could need GPU to increase enormous vector calculations.

[Jouni]

I am really surprised about current tie! SF won last superfinal and according to framework it's improved by 100 points. Lc0 is improved more?

[mwyoung]

I am really surprised about current tie! SF won last superfinal and according to framework it's improved by 100 points. Lc0 is improved more?

You need to think outside of your Elo box. Is it possible for the frame work results to be correct, and the TCEC result. Yes. Testing condition matters.

This result is not surprising at all for testers who tested Stockfish at long and short time controls.