Does Stockfish have a disadvantage when using NNUE in certain scenarios?

[icy kite]

why would you limit depth rather than playing actual games or limiting nodes

[river jackal]

I want to test the performance of NNUE Stockfish against Traditional Stockfish in different scenarios with varying depth limits.

[river jackal]

By the way, "limiting nodes" and "limiting depth" aren't the same thing?

[icy kite]

No

you can tell SF to search a specific number of nodes rather than an arbitrary depth

[river jackal]

Got it, Thank you!!! Gonna fix that.

But that still not explaining how traditional stockfish outperforms nnue stockfish in this specifc scenarios

[icy kite]

would perhaps help to mention what SF version this is

cuz it's definitely not latest

in general though, while NNUE is overwhelmingly stronger than classical eval, it's not perfect

but the strength is the important part

[river jackal]

it is stockfish 16, i'm going to update to stockfish 17 and make the experiment again

[river jackal]

it's not perfect, but should be better than the classical, right? that is my point

[icy kite]

it is

in actual games it's triple digit elo stronger

[river jackal]

Then I must be doing something wrong, I'll review the experiment and bring updates

[icy kite]

not necessarily

these are infinitesimal sample size tests on contrived positions

[modest linden]

limiting nodes might be unfair towards sf classical

[icy kite]

using modern(ish) sf with nnue disabled is unfair towards sf classical

[river jackal]

You knew that it wasn't the latest version because the latest version dont allow me to turn off NNUE? Is it possible to disable NNUE in stockfish 17?

[fresh yoke]

No, HCE no longer exists in SF

[tribal coyote]

Yes, it has disadvantage because the NNUE might not be the best, making the engine moves a bit biased.

[icy kite]

what are you on about

[tribal coyote]

Just telling the truth based on my observation. Stockfish keeps on repeating the same move mistake even in a new game if it explore that particular line.

[icy kite]

idk about "the truth"

you certainly said something

[tribal coyote]

??

[vagrant oak]

?

[modest linden]

i don't see how that is the case

[heavy hawk]

nnue is definitely the best...

[blazing tree]

why not limit by time?

[icy kite]

or nodes

or anything meaningful

[river jackal]

How whould i test the importance of the depth to an evaluation without limiting depth?

[heavy hawk]

limiting by depth is extremely unfair to sf hce. It would effectively have 2-3x less search time

[river jackal]

i'm not limiting time in any scenario

[modest linden]

effectively

[river jackal]

Sorry, what you mean by that? i don't get it.
English is not my native language

[mellow crest]

It would effectively have 2-3x less search time
The "effectively" in that sentence means that limiting the depth like that would have an effect similar to what you'd get by cutting SF HCE's time to half or a third as much.

[cyan spire]

yes, shashchess dev did something like this, i think he keeps updating an older version of stockfish with newer updates with the net removed

[icy kite]

mentioning amchess

[wet coral]

I know stockfish sucks at checkmating with a knight and bishop

[icy kite]

No

[twin oar]

don't know what stockfish you're using then lol