Stockfish now benefits from hyperthreading

[Gusev]

I ran a copy of my in-house Stockfish compile (Stockfish, 16 threads) against its copy (Stockfish_2015-11-03_x64_bmi2_MinGW, 8 threads),
no ponder, no Syzygy, 256MB hash, on a Corei7-5960X (8 cores, hyperthreading on), 20"+0.3" TC, on Marijan's Arena-Blitzer, with the following results

-----------------Stockfish-----------------
Stockfish - Stockfish_2015-11-03_x64_bmi2_MinGW : 509.5/1000 148-129-723 (===0==10=0=01==1====1==0====0====101=0===1====1======0=0=0=0==101=1======01=0=0====1=====1=11=10===0==1=====0========0=1======1001=01=======1===1===01=
==0=01==1==11=========0=========1=0=====0========
=0==0==0==10==111====1=====1===0=======0==0=11==1========0======1=======0==01=10======1========1======0======0======0==========1=11=======0=1======0======1100===0==10===0===0===1=0=1=====0=====1=1=0============1=1==110===1==0===1==1=1===11======1=0=00
=======10===1========0=1==1========1========110==
===0==1===1==1=111==10========1====1====1=1====1==1============0==1=0======1====1===0====0===0======10======0==00===011==0=11=110=========1=======11===01===10======1=1==01=====10=0======0=10==01===1==1==1=====0====1==010==1====110====00======0======0==1====1==1=1===1==0==0===0=====1==1=====1======1===0
1===010=01=10=======0=======0======01=================0==1===1====1====0=====1====0==00========0===0
1===0======0====0====1=1=1=0=========1==0====1==0====0==001====01=0==10==10====0010==1===1=1=1000) 51% +7
-----------------Stockfish_2015-11-03_x64_bmi2_MinGW-----------------
Stockfish_2015-11-03_x64_bmi2_MinGW - Stockfish : 490.5/1000 129-148-723 (===1==01=1=10==0====0==1====1====010=1===0====0======1=1=1=1==010=0======10=1=1====0=====0=00=01===1==0=====1========1=0======0110=10=======0===0===10===1=10==0==00=========1=========0=1=====1========
=1==1==1==01==000====0=====0===1=======1==1=00==0========1======0=======1==10=01======0========0======1======1======1==========0=00=======1=0======1=
=====0011===1==01===1===1===0=1=0=====1=====0=0=1============0=0==001===0==1===0==0=0===00======0=1=11
=======01===0========1=0==0========0========001=====1==0===0==0=000==01========0====0====0=0====0==0============1==0=1======0====0===1====1===1======
01======1==11===100==1=00=001=========0=======00===10===01======0=0==10=====01=1======1=01==10===0==0=
=0=====1====0==101==0====001====11======1======1==0====0==0=0===0==1==1===1=====0==0=====0======0===10===101=10=01=======1=======1======10=============
====1==0===0====0====1=====0====1==11========1===10===1======1====1====0=0=0=1=========0==1====0==1====1==110====10=1==01==01====1101==0===0=0=0111) 49% -7

Is the apparent advantage of 16 threads over 8 due to the benefits of LazySMP?

[AlvaroBegue]

I get a LOS of 0.87319, which means your result is not very significant. A p-value of 0.12 is not good enough even for social science.

[Dann Corbit]

The thing that is very interesting is that it is not clearly worse.

[Gusev]

Whoever is interested is most welcome to run more games. As Dann Corbit pointed out correctly, hyperthreading used to make performance clearly worse in the past, so I am delighted to see that not happen in this one experiment.

[ernest]

Stockfish - Stockfish_2015-11-03_x64_bmi2_MinGW : 509.5/1000 148-129-723

With 95% (2 sigma) error-bar, the result reads as : (509.5 ± 17) / 1000

[Gusev]

As a follow-up, I ran 1000 games of Cheng 4, 8 threads vs. 4 threads on a 4-core i7-930. The 8-thread Cheng 4 won,
-----------------Cheng4_x64-----------------
Cheng4_x64 - Cheng4_x64 - Copy : 523.0/1000 236-190-574 (1=010=1=0101=00=1==01===010==001=10=0==0=====111=11=100=11=1=01=111001011=0===11==01===10======0=110=1==0=00=10====11=0101====011=01=1====1=1=====110=====10====
==1100====1==1==101====1=1=0=1=11==1========0===0====1111=11=00==0==1====101==0=1====1=1====0=00=====11=000========01===10==0=10=====10===0==1==0==00==1=0
=1===1=01=011===111======1011=0===11====0=1=00010=0001==0=1====0100=====0==0=======0==0=========00==11=1=10=0====1===0=1=10110=110==0=100==0=100=01=1=10=0=11=
=====1==1====1=1=0=1=01======1=01=0=0=01=1=1010====1010==1=1=1==001===11=10==101=01=1=01110=01==0===0===1=0=110===0=1011=01==11==01=00===1==1===11=10=11===1===0=10====0==========1=0==0=1========1==========1=1====0===01=011101=1=0=0=1===01010=======11==1==1=1=0==
101=10=0=00=0===1110==0=================1=0=1=1=01=1=1===1======1=====01=11=0===11=11=1=01=0===========
===0111=00===0==01==========1===1=111===01===0===10===1=0101=1=011======0====1===0===100===1=====0=0======0010=1=00====0==110===010=====1==01==1=1=0==1=00====0=001) 52% +14
-----------------Cheng4_x64 - Copy-----------------
Cheng4_x64 - Copy - Cheng4_x64 : 477.0/1000 190-236-574 (0=101=0=1010=11=0==10===101==110=01=1==1=====000=00=011=00=0=10=000110100=1===00==10===01======1=001=0==1=11=01====00=1010====100=10=0====0=0=====001=====01====
==0011====0==0==010====0=0=1=0=00==0========1===1====0000=00=11==1==0====010==1=0====0=0====1=11=====00=111========10===01==1=01=====01===1==0==1==11==0=1
=0===0=10=100===000======0100=1===00====1=0=11101=1110==1=0====1011=====1==1=======1==1=========11==00=0
=01=1====0===1=0=01001=001==1=011==1=011=10=0=01=1=00======0==0====0=0=1=0=10======0=10=1=1=10=0=0101====0101==0=0=0==110===00=01==010=10=0=10001=10==1===1===0=
1=001===1=0100=10==00==10=11===0==0===00=01=00===0===1=01====1==========0=1==1=0========0==========0=0=
===1===10=100010=0=1=1=0===10101=======00==0==0=0=1==
010=01=1=11=1===0001==1=================0=1=0=0=10=0=0===0======0=====10=00=1===00=00=0=10=1==============1000=11===1==10==========0===0=000===10===1===01===0=1010=0=100======1====0===1===011===0=====1=1==
====1101=0=11====1==001===101=====0==10==0=0=1==0=11====1=110) 48% -14

[Gusev]

Likewise, Cheng 4 (16 threads) won over Cheng 4 Copy (8 threads) on i7-5960X,
-----------------Cheng4_x64-----------------
Cheng4_x64 - Cheng4_x64_Copy : 519.0/1000 215-177-608 (1===00=====1==0==0=1====0=01==001=0=010==1=0=1111000=1==00===01==0=10========1=00=01==1=11=0==1==1===110====11=0==0==1=101=100====00===11====11===110===1===
0111==0=1======0=01==1===01====0=11==0=0=0=10===110010===1110==1==0===11==1==101==1==========1=0==1=0===1=====1=1==0=1==011=1==0===1==0=0====1==1==01==0=11=
1=0======111=====00=01=0=1=001=====10==0011==1==0=1======1==0=1=======1===01====1====0=10=011010============0=======0=01===0=01===0===101=======1====0===
==1=1==1=0==1=1====11==010========1==01===10==110010====01===01=====10===0===000=0==1=11===1=00===0==1=1===1=0=111===1======0===00=====01=1==1==10=====0==1
01=0=1=0==110====1====0==0====1===0=1010=1==1====1=1001101===========1=00==1==11==1===010=1======1=1==0=====01===1==========0=1=0=10========0011=010==00==1
=0=0=======1==1===1=1===1=1000==1====0=01==11==1=00==1===1=101==1==0===1====11==11=====0=1===0=0=000==10=10==1===0==1=====0====10==01=1==0==11==101==0=====
=0==0=====01======1=00===0=1=1111===1===0=====0=0=1001==00=1=1011=11=1) 52% +14
-----------------Cheng4_x64_Copy-----------------
Cheng4_x64_Copy - Cheng4_x64 : 481.0/1000 177-215-608 (0===11=====0==1==1=0====1=10==110=1=101==0=1=0000111=0==11===10==1=01========0=11=10==0=00=1==0==0===001====00=1==1==0=010=011====11===00====00===001===0===
1000==1=0======1=10==0===10====1=00==1=1=1=01===001101===0001==0==1===00==0==010==0==========0=1==0=1===0=====0=0==1=0==100=0==1===0==1=1====0==0==10==1=00=
0=1======000=====11=10=1=0=110=====01==1100==0==1=0======0==1=0=======0===10====0====1=01=100101============1=======1=10===1=10===1===010=======0====1===
==0=0==0=1==0=0====00==101========0==10===01==001101====10===10=====01===1===111=1==0=00===0=11===1==0=0===0=1=000===0======1===11=====10=0==0==01=====1==0
10=1=0=1==001====0====1==1====0===1=0101=0==0====0=0110010===========0=11==0==00==0===101=0======0=0==1=====10===0==========1=0=1=01========1100=101==11==0
=1=1=======0==0===0=0===0=0111==0====1=10==00==0=11==0===0=010==0==1===0====00==00=====1=0===1=1=111==01=01==0===1==0=====1====01==10=0==1==00==010==1=====
=1==1=====10======0=11===1=0=0000===0===1=====1=1=0110==11=0=0100=00=0) 48% -14

[Dann Corbit]

This is a very surprising result.
NPS clearly should not be expected to increase.
With double the threads, there should be greatly increased SMP loss.
There must be some other unaccounted affect here.

I would love to see a logical explanation for it.

[AlvaroBegue]

This is a very surprising result.
NPS clearly should not be expected to increase.

Why shouldn't NPS increase? The promise of hyper-threading is that you can effectively run up to two threads per core, with some degradation. This degradation has traditionally been too great to compensate for the sub-linear scaling in parallel-search performance with the number of threads. It could be that recent i7 processors are doing hyper-threading so well that it now does compensate. Or perhaps parallel search now scales better with number of threads?

[Dann Corbit]

Even if the context switch occurs with zero overhead, it is astonishing to imagine that somehow the single core can do more work by simply switching context. That is all that it is doing.