CPU power and latency

Good questions. FWIW I tried a quick-test for the overall latency of my pianoteq-subnotebook. The timespan between the klick-noise of a fingernail hitting a key and the sound from the loudspeaker is around 37 ms as you can see in the picture. This is the most pessimistic scenario, because a midi-note is normally not triggered, when the finger touches the surface, but later, when the key is depressed and reaches the second (or third) keycontact with a specific velocity.

But 37 ms is higher, than I expected, taking into account, that my normal settings are 48000 Hz with buffersize of 64 samples (1.3 ms), which means 2.6 ms latency alone for passing the audio buffer of the onboard PCI-soundcard.

The remaining unknown variables are the latency for processing a tone and the MIDI-signalling via USB from my Kawai to the subnotebook (Celeron N2930).

As far as I remember, real pianos are similar "slow" with around 30 ms.

Thanks for your reply, groovy. Could you by any chance repeat the test using your Kawai's internal sounds, to compare?

Hello snowjoke,

I tested it a long time ago and got 22 ms from a Kawai ES3 alone, see:

http://www.forum-pianoteq.com/viewtopic...8593#p8593

I have an estimation for the "unknown variables" eventually. I don't know, how to measure it exactly, but I compared the note-on-latency of my USB-keyboard with the fastest midi-keyboard I have: the virtual-keyboard inside pianoteq clicked with an USB-mouse!

From the noise of the mouseclick to the sound from the speaker I measured ~19 ms (average of five clicks). Example:

And then I repeated the same with my Kawai ES3 connected to the subnotebook with an USB-cable. The average response from 5 hits (forte) was ~35 ms. Example:

So triggering a midi note-on from my USB-keyboard lasts ~16 ms longer than triggering a note from the virtual-keyboard with my USB-mouse (on my system).

EDIT: Assumed that my overall latency is ~35 ms and ~19 ms  is the best overall latency I can achieve with the virtual-keyboard, then ~16 ms  are needed just to produce the midi-on-event in pianoteq with my real USB-keyboard.

Why is the virtual keyboard ~16 ms faster? Using the real keyboard the (USB) midi-signal has to pass the snd_usb_audio kernelmodule ("driver") on the notebook. Plus the mouse button has a shorter contact-distance than a key. Plus unknown initial latency for midi-note-processing inside the Kawai.

The only parameter that could be influenced is the performance of the snd_usb_audio "driver". With older 2.6x-linuxkernels there was an option 'nrpacks' for example, that could be reduced from 8 to 1, and this shorter chunks improved USB-latency. In newer kernels, like my 3.16x from this experiment, nrpacks is obsolete now and handled automatically AFAIK. BTW, does someone know, if USB-latency can be tweaked somehow nowadays in 3.x-kernels?

cheers

PS: Keep in mind, that the overall latency of course is influenced by the speed of the keypress. Hitting my test-key with ff I saw latencies down to 30 ms (rather than ~35 ms)

AFAIK, an SSD would not have any influence with PianoTeq: everything is done in RAM.

Assuming JACK audio server daemon is running with these options:

'jackd -dalsa -p128 -n3 -r44100'

(frames/period)*(Periods/buffer)/(sample rate)

128 * 3 / 44100/s = 8.7 ms

Is the calculated "latency" of 8.7 ms correct then?

Is it the latency between the modelled tone and its entry into the audio-hardware? In other words, is 8.7 ms the time spent in JACK audio server until a sample from pianoteq reaches the audio card?

Thanks

There are command line utilities to control CPU like cpufreq.   Probably there are an applet you can use in gnome3.   Google is your friend.

groovy - thanks for posting your interesting findings. 16 ms seems like a large difference but I suspect it's mainly due to the time taken for the key to move. Maybe I will do my own experiment to see if I get similar results (using Windows).

varpa - thanks for your on-topic reply! So, buying a faster computer won't improve my latency, if my current machine can cope with the lowest buffer settings already.

When you say "Increasing latency will reduce the CPU load", do you mean that processing of particularly intense moments can be spread out, hence peak load will be lower? The CPU needs to do the same amount of work whatever the latency, and you already said that latency doesn't depend on CPU power.