SMP SF Formula

[Werewolf]

My point is this:

Is the issue that multi-threading somehow doesn't work well with NNUE, or is it that you get exactly the same phenomena when time is greatly increased as well?

I'm wondering if NNUE gave such a large gain that increasing time per move (or processing speed) yields less gain than it used to.

[smatovic]

SF uses an own implementation of LazySMP parallel search, it is not comparable to things you can figure out on paper like DTS, YBWC or ABDADA, therefore they measure Elo gain and not time to depth speedups, I guess SF SMP is intertwined heavily with their selective search heuristics, some say they just widen the search with their approach, but IMO it is not that simple.

--
Srdja

[Jouni]

I have no reason to doubt Young results. But he uses Perfect Book 2021, which have obviously very drawish lines.

[Werewolf]

SF uses an own implementation of LazySMP parallel search, it is not comparable to things you can figure out on paper like DTS, YBWC or ABDADA, therefore they measure Elo gain and not time to depth speedups, I guess SF SMP is intertwined heavily with their selective search heuristics, some say they just widen the search with their approach, but IMO it is not that simple.

--
Srdja

Understood.
But if the SMP search has much lower gains than before as the thread count increases, it does suggest something is wrong.

[lkaufman]

SF uses an own implementation of LazySMP parallel search, it is not comparable to things you can figure out on paper like DTS, YBWC or ABDADA, therefore they measure Elo gain and not time to depth speedups, I guess SF SMP is intertwined heavily with their selective search heuristics, some say they just widen the search with their approach, but IMO it is not that simple.

--
Srdja

Understood.
But if the SMP search has much lower gains than before as the thread count increases, it does suggest something is wrong.

Elo gains will be lower due to higher strength. The proper test here would be to have the NNUE version on eight threads play against the same version on one thread with 5x as much time, then repeat with NNUE turned off on both. Both results should be close to 50%, if one does much better than the other it indicates it is more MP friendly.

[Werewolf]

SF uses an own implementation of LazySMP parallel search, it is not comparable to things you can figure out on paper like DTS, YBWC or ABDADA, therefore they measure Elo gain and not time to depth speedups, I guess SF SMP is intertwined heavily with their selective search heuristics, some say they just widen the search with their approach, but IMO it is not that simple.

--
Srdja

Understood.
But if the SMP search has much lower gains than before as the thread count increases, it does suggest something is wrong.

Elo gains will be lower due to higher strength. The proper test here would be to have the NNUE version on eight threads play against the same version on one thread with 5x as much time, then repeat with NNUE turned off on both. Both results should be close to 50%, if one does much better than the other it indicates it is more MP friendly.

Yes but if that's true you'd get exactly the same phenomena observed from large increases in time as well. If going from 1 thread to 16 only yields a small gain now, then a large time increase (say 10x) would be similar.

It's not an SMP issue actually if your point stands.

[MikeB]

https://github.com/official-stockfish/S ... ssues/3727

[Werewolf]

Very informative post.

It's still not clear in my mind whether the lack of elo gain going from 1 thread to many is caused by:

a) Engines just getting stronger, so the draw ratio increases.
b) An issue with SMP with NNUE and point a) as well.

If it's a) you should see the same phenomenon when two identical engines with equal threads have a massive time discrepancy. If it's b) you'll only see this phenomenon mildly with such conditions.

[Raphexon]

Very informative post.

It's still not clear in my mind whether the lack of elo gain going from 1 thread to many is caused by:

a) Engines just getting stronger, so the draw ratio increases.
b) An issue with SMP with NNUE and point a) as well.

If it's a) you should see the same phenomenon when two identical engines with equal threads have a massive time discrepancy. If it's b) you'll only see this phenomenon mildly with such conditions.

What issue with SMP?

[Werewolf]

A short match, but something nevertheless:

Engine: Stockfish 14.1
Book: Nooman short
TC: G10 + 10
Threadripper 3990X
60 threads vs 1 thread
Ponder = on (keeps clockspeed fair)

60 draws
6 wins for the 60 thread version
0 Losses

54% or +30 elo for 60 threads vs 1 thread.