Board representation, evaluation speed and NNUE

[gaard]

My understanding is that bitboards make conventional position evaluations significantly faster, especially compared to mailbox, and that was primarily why many people abandoned mailbox years ago. Does NNUE negate that advantage any or is NNUE performance also heavily dependent on board representation?

[AndrewGrant]

My understanding is that bitboards make conventional position evaluations significantly faster, especially compared to mailbox, and that was primarily why many people abandoned mailbox years ago. Does NNUE negate that advantage any or is NNUE performance also heavily dependent on board representation?

NNUE should not be impacted very much by board representation. Most updates are incremental, and I can see it being of equal cost both ways. The performance gap in NNUE is Fused-Multiply-Adds from the Input Layer to Layer 1, not the way you define the board structure.

[gaard]

My understanding is that bitboards make conventional position evaluations significantly faster, especially compared to mailbox, and that was primarily why many people abandoned mailbox years ago. Does NNUE negate that advantage any or is NNUE performance also heavily dependent on board representation?

NNUE should not be impacted very much by board representation. Most updates are incremental, and I can see it being of equal cost both ways. The performance gap in NNUE is Fused-Multiply-Adds from the Input Layer to Layer 1, not the way you define the board structure.

In your opinion, then, does this negate the main performance advantages of a bitboard engine?

[Sopel]

It doesn't negate it, just makes it relatively smaller

[AndrewGrant]

It doesn't negate it, just makes it relatively smaller

+1. I would still use bitboards though.

[amanjpro]

If you can make your move generator as fast, and SEE computation as fast using another techniques other than bitboards, then yes.

But I'm my very little experience bitboards shine in these two areas more than any other technologies

[hgm]

A good mailbox implementation is faster than bitboards. But with NNUE the evaluation gets to dominate the time usage completely. And unlike HCE, it doesn't take any advantage from one representation or another.

[dangi12012]

A good mailbox implementation is faster than bitboards. But with NNUE the evaluation gets to dominate the time usage completely. And unlike HCE, it doesn't take any advantage from one representation or another.

I can 100% confirm that this is not the case at all. There are a list of about a dozen optimisations that are not applicable to mailslots but to bitboards only.
Btw I can say that for sure because I implemented dozens of different approaches and it always goes that all mailslot ideas can be implemented in bitboards but not always the other way around:

Movegen - Single Thread - No Hash - Bulk counting Enabled.
Perft 7: 3195901860 2180ms
Perft Kiwi 6: 8031647685 3921ms

But yeah CPU time will be spent in NNUE - or pruning + sorting in general. Since every move not saved will translate to literally billions of wasted moves a few steps down the tree.

[AndrewGrant]

A good mailbox implementation is faster than bitboards. But with NNUE the evaluation gets to dominate the time usage completely. And unlike HCE, it doesn't take any advantage from one representation or another.

I can 100% confirm that this is not the case at all. There are a list of about a dozen optimisations that are not applicable to mailslots but to bitboards only.
Btw I can say that for sure because I implemented dozens of different approaches and it always goes that all mailslot ideas can be implemented in bitboards but not always the other way around:

Movegen - Single Thread - No Hash - Bulk counting Enabled.
Perft 7: 3195901860 2180ms
Perft Kiwi 6: 8031647685 3921ms

But yeah CPU time will be spent in NNUE - or pruning + sorting in general. Since every move not saved will translate to literally billions of wasted moves a few steps down the tree.

Bulk counting would obviously favor bitboards. General case, I think they are not so different speed wise. Both have advantages.

[hgm]

Perft speed is meaningless for an engine. Engines do so many things that are ignored in perft. And they must mainly generate captures.