Stockfish with new NNUE architecture and bigger net released

[glennsamuel32]

Oh, the other famous and successful double sized net.
What was that again?

I resign as moderator from this forum.
It's time for me to take a break.
Maybe some photography.

Or perhaps fishing
You will have a.....stock of fish !

May I suggest cast "net" fishing instead ?
The 512 series appear to catch the most !!

[Rebel]

A moderator who resigns is a serious matter, no more jokes please.

[mclane]

...
You're right. This new try does not seem nearly so successful.

Surely you can't be serious, unless you're defining "success" in some nonsensical way. I don't see how any rational person could come to your conclusion given the facts. Hopefully you're confused rather than arguing in bad faith. This network is almost certainly stronger than the FF2 network. In fact, the FF2 network proved inferior to even the SF13 network H2H in all statistically significant independent tests I've seen. The FF2 network wouldn't have even passed FishTest. I think you pointed to CEGT showing FF2 ahead of SF 13 earlier, but the relative elo is well within error bars (NVM the fact that SF13 and SF12 won H2H there too, though the sample size is tiny).

As many others have pointed out and you're neglecting to acknowledge, Albert wasn't the first to try training wider networks. He didn't even succeed in demonstrating the superiority of larger networks. Albert trained a fairly strong network using the available tools, but this likely isn't even close to as difficult as you're imagining.

Of course, I am trolling. But I am also trying to make a point. I think that Albert is being vilified in an unfair and (dare I say silly?) way.

People are mad because he made false claims, tried to paint himself as an innovator and didn't abide by the GPL license.
He also made the claim of doing it on his own while also having asked SF-devs for help numerous times. If any real innovation happened, it's his programmer friends that did it. (Daniel Uranga and Dkappe)

this was the first public example of the NNUE net attempted at double size which makes it both clever and interesting.

It wasn't.

Oh, the other famous and successful double sized net.
What was that again?

I resign as moderator from this forum.
It's time for me to take a break.
Maybe some photography.

Why do you find it ok to make money out of a gpl licence ?!

If something is for free and somebody puts a label on it : fat, this makes it a valubale product ???

[jr66]

A moderator who resigns is a serious matter, no more jokes please.

I thought he was not serious ???
If so, my apologies !

[Cornfed]

Why do you find it ok to make money out of a gpl licence ?!

If something is for free and somebody puts a label on it : fat, this makes it a valubale product ???

You did forget that it was bundled with the GUI more people probably have than any other...
And he doubled the size and had a minor tweak here or there.

Let the market decide...if people needed a new GUI and that persuaded them to buy it (they could of course run any 'ol Stockfish under the current or earlier Fritz...) then that's their decision.

[ChickenLogic]

Oh, the other famous and successful double sized net.
What was that again?

I showed that "double sized nets" are much stronger on a node per node basis months before nnue was merged and Albert knew about nnue. I also showed that, at the time I trained those nets, the slowdown was too severe for the net to make up for it. Jjosh basically did the same. This was due to a poorly optimized feature transformer etc. The "innovation" was already there and we were open about it and I specifically told Albert about this. This is why it's outrageous to attribute it to him. Just because he has friends in a well known company (at least in the chess world) and "his" net thus became "famous" doesn't mean he's the inventor or innovator of it. I don't get your trouble in understanding this.
Moreover, the nets internal structure that's independent of size is what makes the net unique. Size can be altered at any given time and it's done all the time to find if there's a new optimal size. That alone disqualifies a "new size" from being an "innovation".

The real innovation was to optimize the feature transformer and other parts of the code to the point the slowdown lost less elo than the net gained in accuracy. None of these were written by Silver.
This should be obvious to anybody who looks at how the code evolved and the discussions and idea exchanges in the Stockfish discord/github/fishcooking.

Just on a side note: just today I trained a net stronger than FatFritz2 in under a day with maybe 30 minutes of "actual labor" + some electricity which comes for free as they stem from solar panels. And I didn't rent massive hardware either, like Albert did, I "merely" used a 980TI. This should give people an idea on how comparatively easy it is to train a strong nnue net. With a RTX 2060 one should be easily able to get within a few elo of SF master in a day.

On another side note Dann, taking a break from any hobby can sometimes be a real relief. Pursuing new things is refreshing and leaving old hobbies "behind" is not necessarily a bad thing. If you decide to take on photography I wish you all the best. I think you're a good person - I just disagree heavily with you on this one topic and that's perfectly fine.

[AlexChess]

test results against prior version of SF:

https://tests.stockfishchess.org/tests/ ... 3412d0921d
TC: 10s+0.1s, 1 thread
ELO: 21.74 +-3.4 (95%) LOS: 100.0%
Total: 10000 W: 1559 L: 934 D: 7507 Elo +21.74
Ptnml(0-2): 38, 701, 2972, 1176, 113

https://tests.stockfishchess.org/tests/ ... 3412d0925b
TC: 60s+0.6s, 1 thread
ELO: 5.85 +-1.7 (95%) LOS: 100.0%
Total: 20000 W: 1381 L: 1044 D: 17575 Elo +5.85
Ptnml(0-2): 27, 885, 7864, 1172, 52

https://tests.stockfishchess.org/tests/ ... 7940a03806
TC: 20s+0.2s, 8 threads
LLR: 2.93 (-2.94,2.94) <0.50,3.50>
Total: 34272 W: 1610 L: 1452 D: 31210 Elo +1.60
Ptnml(0-2): 30, 1285, 14350, 1439, 32

https://tests.stockfishchess.org/tests/ ... 7940a03c72
TC: 60s+0.6s, 8 threads
LLR: 2.94 (-2.94,2.94) <-2.50,0.50>
Total: 45544 W: 1262 L: 1214 D: 43068 Elo +0.37
Ptnml(0-2): 12, 1129, 20442, 1177, 12

Hi MikeB !

On my test (Blitz 3 mins + 3 secs /all Arena Windows 10 4CPUs) Honey 13.2 with new 40 MB NNUE nn-7756374aaed3.nnue is much weaker than 13.1 (20th 17.5 vs 7th 21.0 points)



Regards, AlexChess

[AlexChess]

I'm not sure that the new default NNUEs

21-05-20 20:30:09 nn-7756374aaed3.nnue AND 21-05-16 17:39:14 nn-8a08400ed089.nnue by vdv

are the best performers on my slow system NOT AVX2 | BM2.

21-05-06 09:15:53 nn-82a4dd55cc6d.nnue by Sopel 47.7 MB is giving me better results with latest Stockfish and derivates
(x64 4CPUs 3-4 mN/s)

Regards, AlexChess

[MikeB]

I'm not sure that the new default NNUEs

21-05-20 20:30:09 nn-7756374aaed3.nnue AND 21-05-16 17:39:14 nn-8a08400ed089.nnue by vdv

are the best performers on my slow system NOT AVX2 | BM2.

21-05-06 09:15:53 nn-82a4dd55cc6d.nnue by Sopel 47.7 MB is giving me better results with latest Stockfish and derivates
(x64 4CPUs 3-4 mN/s)

Regards, AlexChess

Yes, it is clear that old systems are setting the short end of the stick with code that is taking advantage of moremodern hardware.
Regardess, still unbeatable by most humans at full strength ;>)

[Cornfed]

Yes, it is clear that old systems are setting the short end of the stick with code that is taking advantage of moremodern hardware.
Regardess, still unbeatable by most humans at full strength ;>)
[/quote]

"Most"? Dare I say 'any'?