Stockfish now benefits from hyperthreading

[bob]

Thank you!! Let's not take the current, limited results too seriously. Which chess interface do you usually use for testing?

I use one that I wrote that is VERY lightweight with no GUI or anything. Works best on my cluster stuff...

[Gusev]

As the next follow-up, I ran a 2000-game match of Cheng 4, 16 threads vs. 8 threads, on the 8-core i7-5960X, at the TC of 1'+1", 256MB hash, HT on. The result is similar to those seen before, 16 threads won, Cheng4_x64 - Cheng4_x64_Copy : 1042.0/2000 419-335-1246, 52% +14
I have also purchased a copy of Komodo 9.3, so the next match will test how well it handles hyperthreading.

[Gusev]

16 threads win over 8 threads with the newly purchased Komodo 9.3 on the 8-core i7-5960X, at the TC of 1'+1", 256MB hash, HT on:
Komodo-9.3-64bit - Komodo-9.3-64bit - Copy : 1063.5/2000 312-185-1503 53% +21
I will post the result of the 8 threads vs. 4 threads matchup on a quad next.

At 2000 games, unless there is a procedural error of some sort, this is enough games to show that hyperthreading is probably not a detriment for Lazy SMP.

In itself, that is incredibly surprising.
As we add more and more threads, we should have more and more SMP loss.
But we don't.

I would love to see a logical explanation for that.
I can't think of any.

[Gusev]

On a Core i7-930 quad, 8 threads beat 4 threads, Komodo-9.3-64bit - Komodo-9.3-64bit - Copy : 1095.0/2000 391-201-1408, 55% +35

As the next follow-up, I ran a 2000-game match of Cheng 4, 16 threads vs. 8 threads, on the 8-core i7-5960X, at the TC of 1'+1", 256MB hash, HT on. The result is similar to those seen before, 16 threads won, Cheng4_x64 - Cheng4_x64_Copy : 1042.0/2000 419-335-1246, 52% +14
I have also purchased a copy of Komodo 9.3, so the next match will test how well it handles hyperthreading.