{Not good at stats, but I have a grasp of it}

I was one of the first, if not the first, to imagine and suggest to use non-real time calculation (audio export for example) to make pianoteq use a lot more calculation, more precision, to create perfect piano sounds up to the point to shut-up any critic. Studios, for example, could play the piano in real time( while the artist performs) and later render it using algorithms that requires heavier computation to produce perfect piano sounds. Even if took 1 hour to render 5 minutes of performance, it would be worthwhile.

The problem it's the high cost to redesign the software core sound generator. Other problem is that people could get used with the perfect sound rendered and start to spot/point more deffects in the real time playing mode. But today pianoteq 6.5 it's already very good, close to the real thing.

https://www.amazon.com/Intel-i9-9980XE-...amp;sr=8-7

18 cores of 4,4ghz each...
36 threads...

They would be able compute much more details for ultra hyper realistic sounds.

According to what else was running and what else had run recently - the first measurement could have been slowed by other tasks/processes and or their residue.
The second experiment started on a relatively "clean" machine.

I would therefore do both experiments several times on clean re-boots.
There is still likely to be variation, something you can ponder.

So with hyperthreading turned ON the two lowest latency measurements were 13 and 14 ?
with it OFF they were 13 and 16 ?

this is not a thread about latency in general, there are already a few that go more in deep.

It's just about a single parameter: hyperthreading.

ceteris paribus

Cheers

Thus, it is better (or was at that time) to have two fast cores than four slow ones.

Better in respect of real-world latency? That's the question.

Minimum latency does not necessarily reflect some Benchmarking indices.

What has better latency? Two native CPUs or two CPUs emulating four (slower?) virtual CPUs aka "Simultaneous MultiThreading"?

An OpenBSD maintainer (following my upper link) indicated, that the additional multithreading layer could have negative impact on system "performance".

So I made my real-world experiment and found no difference in latency using 2 cores or 2 cores with SMT (= 4 "virtual" CPUs). But this is just a sample.

Edit: this is the post https://www.forum-pianoteq.com/viewtopic.php?id=4149

This question came up recently in Silent laptop advice.

I tried to get an impression with the following quick test:

To get a value for an overall-latency I clicked PTQ's virtual keyboard with a mouse and recorded the click-noise and the PC-speaker sound with the internal microphone of that notebook (velocity fixed at 100).

Then I switched off hyperthreading temporarily by starting the kernel with the option "nosmt", which stands for no simultaneous multithreading, and recorded the same PTQ sound again.

The timespan between the mouseclick and the Pianosound measured with audacity is the latency (example screenshot below).

With hyperthreading on (smt) I got an average overall-latency of 17.1 ms (19, 14, 17, 17, 15, 22, 20, 13).

With hyperthreadinng turned off (nosmt) the average was 18.9 ms (18, 16, 22, 13, 16, 22, 19, 20, 24).

No significant difference of latency found in this experiment.

Some technical details just if interested:

With smt:
---------
$ lscpu
Architecture:          x86_64
CPU op-mode(s):        32-bit, 64-bit
Byte Order:            Little Endian
CPU(s):                4
On-line CPU(s) list:   0-3
Thread(s) per core:    2
Core(s) per socket:    2
Socket(s):             1
NUMA node(s):          1
Vendor ID:             GenuineIntel
CPU family:            6
Model:                 78
Model name:            Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz
Stepping:              3
CPU MHz:               2799.902
CPU max MHz:           2800.0000
CPU min MHz:           400.0000
BogoMIPS:              4800.00
Virtualization:        VT-x
L1d cache:             32K
L1i cache:             32K
L2 cache:              256K
L3 cache:              3072K
NUMA node0 CPU(s):     0-3

Without smt ("nosmt"):
----------------------
$ lscpu
Architecture:          x86_64
CPU op-mode(s):        32-bit, 64-bit
Byte Order:            Little Endian
CPU(s):                4
On-line CPU(s) list:   0,1
Off-line CPU(s) list:  2,3
Thread(s) per core:    1
Core(s) per socket:    2
Socket(s):             1
NUMA node(s):          1
Vendor ID:             GenuineIntel
CPU family:            6
Model:                 78
Model name:            Intel(R) Core(TM) i5-6200U CPU @ 2.30GHz
Stepping:              3
CPU MHz:               2800.048
CPU max MHz:           2800.0000
CPU min MHz:           400.0000
BogoMIPS:              4800.00
Virtualization:        VT-x
L1d cache:             32K
L1i cache:             32K
L2 cache:              256K
L3 cache:              3072K
NUMA node0 CPU(s):     0,1

It is an office-notebook equipped with Pianoteq trial v6.5.3 just for this test. Only optimisation was setting the scaling_governor to "performance" on each CPU.
Testsound was Steinway D Prelude at 44.1 kHz and 64 samples buffersize for better latency, than the default 512 (!) samples. OS Debian Linux "Stretch".

Example latency measured in the wave editor: