Stockfish NNUE Questions

[Stephen Ham]

Hello All,

Question #1:
In an earlier thread today, a member wrote that Stockfish NNUE switches to classical Stockfish (SF) when the evaluation exceeds 2.50. I wondered where that threshold was, so thank you! Regardless, alpha-beta engines are superior in endgames while NN engines have performed poorly there. So, does that mean that SF NNUE will continue to use NN play into endgames as long as the evaluation threshold remains below 2.50? Hopefully the engine is instead coded to switch to classical SF when an endgame is reached, perhaps defined by a preset number of pieces/pawns on the board.

Question #2: The abrok website provides development upgrades for specific PCs. My PC is a AMD Ryzen 9 3900X 12-core, 3.80 GHz, on Windows 10. So, which of the below abrok PC options should I upgrade to for maximum performance? To date, I've always just selected the modern computers option.

Windows x64 for modern computers + AVX2
Windows x64 for modern computers
Windows x64 + SSSE3

Thanks in advance for educating me.

Sincerely,
-Steve-

[Werewolf]

Hello All,

Question #1:
In an earlier thread today, a member wrote that Stockfish NNUE switches to classical Stockfish (SF) when the evaluation exceeds 2.50. I wondered where that threshold was, so thank you! Regardless, alpha-beta engines are superior in endgames while NN engines have performed poorly there. So, does that mean that SF NNUE will continue to use NN play into endgames as long as the evaluation threshold remains below 2.50? Hopefully the engine is instead coded to switch to classical SF when an endgame is reached, perhaps defined by a preset number of pieces/pawns on the board.

So just on your first question, unless I'm really getting this wrong SF NNUE is an alpha-beta all the time. The alpha-beta part is the search method.

What changes at 2.5 pawns is the evaluation function.

[syzygy]

Hello All,

Question #1:
In an earlier thread today, a member wrote that Stockfish NNUE switches to classical Stockfish (SF) when the evaluation exceeds 2.50. I wondered where that threshold was, so thank you! Regardless, alpha-beta engines are superior in endgames while NN engines have performed poorly there. So, does that mean that SF NNUE will continue to use NN play into endgames as long as the evaluation threshold remains below 2.50? Hopefully the engine is instead coded to switch to classical SF when an endgame is reached, perhaps defined by a preset number of pieces/pawns on the board.

So just on your first question, unless I'm really getting this wrong SF NNUE is an alpha-beta all the time. The alpha-beta part is the search method.

What changes at 2.5 pawns is the evaluation function.

Indeed. As far as I know, it's not so much that Lc0's evaluation is bad in the endgame. It's more that Lc0's search cannot compete with SF's deep search in positions where that deep search suffices to solve the position.

[Stephen Ham]

Thank you, gentlemen! That's good to know.

Anyone have any input on Question #2? Although I'm an ICCF GM, I'm a computer dummy. So, I value what I can learn from you experts at this website.

And since I'm showing my computer ignorance here, please allow a third question, which I hope isn't too stupid.
Question #3:
It's my understanding that SF NNUE test nets are developed from self-play. In short, it's SF playing against itself. But that seems one-dimensional as the nets then only get really proficient at defeating SF. Why not also train the nets against Leela and other top NN engines? Learning how to defeat them too should produce the all-around strongest engine, rather than one that's just great at defeating SF. Am I missing something?

Thanks in advance.

All the best,
-Steve-

[phhnguyen]

Thank you, gentlemen! That's good to know.

Anyone have any input on Question #2? Although I'm an ICCF GM, I'm a computer dummy. So, I value what I can learn from you experts at this website.

I don't have a computer similar to yours, thus I don't know which one is the best performance. However, I know the solution you may do yourselve and post the result here to help other people:
- Download all (3) those binaries
- Download the default net from Fishtest and store with those binaries
- Run benchmark for all binaries, for example:

Code: Select all

stockfish.exe bench 16 1 13 default depth NNUE

Delete any crashed one and compare their results, keep the best one. You may run several times to confirm results. I guest all take about 10 minutes

And since I'm showing my computer ignorance here, please allow a third question, which I hope isn't too stupid.
Question #3:
It's my understanding that SF NNUE test nets are developed from self-play. In short, it's SF playing against itself. But that seems one-dimensional as the nets then only get really proficient at defeating SF. Why not also train the nets against Leela and other top NN engines? Learning how to defeat them too should produce the all-around strongest engine, rather than one that's just great at defeating SF. Am I missing something?

Thanks in advance.

All the best,
-Steve-

As far as I know, NNUE evaluation is trained to replace and improve the clasical one. It is not trained to win other engines. The high optimisation of SF search function as well as hybrid mode makes that way can work smothly but it may not work with too-much-different evaluation.

[corres]

I have a Ryzen 9 3950x CPU and I found the most faster SF+NNUE binary (.exe) is the AVX2 optimized one.
It is faster about +20 % than the popc.
But the proposed test will decide the question.

[Stephen Ham]

Thank you gentlemen for your informed thoughts and suggestions.

All the best,
-Steve-

[Modern Times]

I have a Ryzen 9 3950x CPU and I found the most faster SF+NNUE binary (.exe) is the AVX2 optimized one.

Yes, same with my Ryzen 7 laptop,

[zullil]

I have a Ryzen 9 3950x CPU and I found the most faster SF+NNUE binary (.exe) is the AVX2 optimized one.

Yes, same with my Ryzen 7 laptop,

I suppose this is because the bmi2 binaries include the PEXT instruction, which is very slow on Ryzen?

[Dann Corbit]

Yes, you can look at Ipman's chess site, and he has lots of samples where the modern compile smokes the PEXT builds on modern AMD stuff. They emulated PEXT in software and did a bad job of it. It's not really a hardware instruction, even though they call it one