6 cores vs 1 core surprised me

[Jouni]

Because playing takes too much time I tried Arasan test suite with 6 and 1 cores. 200 positions checked to be correct. With 6 cores 15 sec limit, but with 1 core 90 sec. So similar CPU time. Results:
Stockfish17 6 cores 181 and 1 core 189
Obsidian14 6 cores 177 and 1 core 184

[Paloma]

A surprising result.
Maybe the hashtables slow down the 6 cores?

Very good investigation!

[Steve Maughan]

Hi Jouni,

Isn't this what you'd expect? If I've understood correctly, you're giving the 1 core run six times the amount of time to solve the problems than the six core. Multi-core engines get (severely) diminishing marginal benefit from additional cores, so one would expect the single-core engine to out-perform the six-core engine if time per cores is equal.

— Steve

Because playing takes too much time I tried Arasan test suite with 6 and 1 cores. 200 positions checked to be correct. With 6 cores 15 sec limit, but with 1 core 90 sec. So similar CPU time. Results:
Stockfish17 6 cores 181 and 1 core 189
Obsidian14 6 cores 177 and 1 core 184

[Werewolf]

Hi Jouni,

Isn't this what you'd expect? If I've understood correctly, you're giving the 1 core run six times the amount of time to solve the problems than the six core. Multi-core engines get (severely) diminishing marginal benefit from additional cores, so one would expect the single-core engine to out-perform the six-core engine if time per cores is equal.

— Steve

Because playing takes too much time I tried Arasan test suite with 6 and 1 cores. 200 positions checked to be correct. With 6 cores 15 sec limit, but with 1 core 90 sec. So similar CPU time. Results:
Stockfish17 6 cores 181 and 1 core 189
Obsidian14 6 cores 177 and 1 core 184

Exactly. I'm surprised people are surprised.

[Vinvin]

Yes, if I remember well, 4 CPUs system is about the same strength as 1 CPU speed x3 or 3 times longer thinking (without permanent brain). So, for 6 CPUs, my guess is around 4 times longer thinking with 1 CPU.

Hi Jouni,

Isn't this what you'd expect? If I've understood correctly, you're giving the 1 core run six times the amount of time to solve the problems than the six core. Multi-core engines get (severely) diminishing marginal benefit from additional cores, so one would expect the single-core engine to out-perform the six-core engine if time per cores is equal.

— Steve

Because playing takes too much time I tried Arasan test suite with 6 and 1 cores. 200 positions checked to be correct. With 6 cores 15 sec limit, but with 1 core 90 sec. So similar CPU time. Results:
Stockfish17 6 cores 181 and 1 core 189
Obsidian14 6 cores 177 and 1 core 184

[Werewolf]

Yes, if I remember well, 4 CPUs system is about the same strength as 1 CPU speed x3 or 3 times longer thinking (without permanent brain). So, for 6 CPUs, my guess is around 4 times longer thinking with 1 CPU.

Hi Jouni,

Isn't this what you'd expect? If I've understood correctly, you're giving the 1 core run six times the amount of time to solve the problems than the six core. Multi-core engines get (severely) diminishing marginal benefit from additional cores, so one would expect the single-core engine to out-perform the six-core engine if time per cores is equal.

— Steve

Because playing takes too much time I tried Arasan test suite with 6 and 1 cores. 200 positions checked to be correct. With 6 cores 15 sec limit, but with 1 core 90 sec. So similar CPU time. Results:
Stockfish17 6 cores 181 and 1 core 189
Obsidian14 6 cores 177 and 1 core 184

Are you sure you're not remembering YBWC? In Rybka days - pre Lazy - the formula gain for cores was CORES^0.76
So 6 cores would be 3.9x faster than one core.

I would hope Lazy SMP would beat that, certainly as core counts rise, Lazy would win.

[JohnWoe]

There's nothing surprising since in multi threaded search. The extra cores just populate hash tables which helps the main thread to see deeper.
You just launch extra objects to search the tree. You will get slowdown since you need to avoid corrupting data via various locking mechanisms etc.

[Ciekce]

There's nothing surprising since in multi threaded search. The extra cores just populate hash tables which helps the main thread to see deeper.

wider, rather - or "better", most accurately

You will get slowdown since you need to avoid corrupting data via various locking mechanisms etc.

in practice no one actually does any locking

[Jouni]

Stockfish wiki says threading efficiency is 82-83%. But tested before NNUE. Maybe this confirms just that?

[Ciekce]

Maybe this confirms just that?

it confirms literally nothing, you are looking at test positions

may the gods one day grant you understanding of proper esting, as has been explained to you in any number of fora infinite times