August 21, 2018, 12:07:57 am
Welcome, Guest. Please login or register.
Did you miss your activation email
News:

Arturia Forums



Author Topic: Experiencing performances issues with polyphonic presets modes on low lat system  (Read 486 times)

Humanizer

  • Apprentice
  • *
  • Posts: 18
  • Karma: 0
hello, thanx for Arturia Amazing sounding works.

experiencing crack and pops on presets where polyphony set on 3 and 4. only one instance.

the strange thing is the fact that i can open a least 10 and more instances of the EASEL plugin playing together without any dropouts, at the only condition to stay in monophonic mode for each one.

buffers are set at the minimal (128samples) latency on a optional USB card in LYNX AURORAs converters. Sampling rate is 96kHz.
setup here is running on NUENDO 8.2 in 64bit mode, hosted by WINDOWS 10 pro computer powered by a intel i9 7900X water-cooled cpu running on all cores at 4,7Ghz with 32gigs of RAM and NVME SSD. graphics are handled by two old 980ti in SLI mode.

So seems that there we need some improvements and CPU optimisation here.

Note:NUENDO has been brided to use only 14 logic cores due to the windows 10 mmcss  limitation issue.see more info here:
https://helpcenter.steinberg.de/hc/fr/articles/115000535804-Windows-10-audio-dropouts-on-multi-core-CPU-setups

Regards.

EDIT:this Behavior seems hit other Arturias plugs : MATRIX12V2 is impacted same manner: POLYPHONY is dramatically impacting CPU perfs too !!!!
CPU is much more impacted by 1 instance of 12 voices than 12 instances of 1 voice!!! is there a reason...? ???

EDIT 2: Modular V2 impacted too: impossible to play a 32 voice unison preset BUT i can easily open and realtime play 50 different instances of Modular without a blink of the system.....
SEM V doesn't seems to have this behavior.....

and what about Arturia other products??? please investigate.


« Last Edit: May 28, 2018, 07:10:24 pm by Humanizer »

LBH

  • Hero Member
  • *****
  • Posts: 1.681
  • Karma: 52
Most DAWs and other music software still put most of the load on a single CPU core for a track or process. When you use multiple instances, then more CPU cores are used.
Check if you are overloading a single CPU core.
The more voices used in a sound, the more demanding it is.

A buffer at 128 samples on 96000 Hz is quite demanding for your CPU. Have you tried for example 44100 or 48000 Hz - or for example a buffer at 256 samples? Perhaps that can give some answers.

Thanks for the Steinberg link. Interesting reading. I have read about more changes with windows 10, and i think developers need to change some thing in the coding. But in this case, then this might not be the issue, as you are running at demanding settings. Also be sure, that your PC is set up to perform it's best for your needs.

BTW: Clicks and pops issues have been reported.Perhaps it's worth to try out, if polyphony is a cause - just in case.

Humanizer

  • Apprentice
  • *
  • Posts: 18
  • Karma: 0
hello again
there are some screenshots of few bench i had done last night:

https://www.dropbox.com/sh/scj1tx33fqee0nd/AADKeJuv4JbVxm2zMI6b4RXZa?dl=0

note the differences when monitoring is ON or OFF (Steinberg ASIOGUARD systems plays a role).it is not an issue.

note the terrible CPU impact load when the POLYPHONY is set to 4 on only one instance. it is  impacting not only one core but ALLs logic cores used by NUENDO.

there obviously a big performance problem on theses puppys. is it related to the intel I9 architecture? or NUENDO STEINBERG mechanism? or WIN 10 behavior? i will let Arturia guess.

please note that have minimum latencies and 96kHz is a real bargain. Almost all of plugins i use sound better in 96kHz. So forget  lower the resolution to get rid of that problem.

As i get VERY GOOD PERFS with others brand of VSTs, pc is good, hi end CPU, Flagship Steinberg DAW, updated operating system and updated Arturia's(not cheap).  i expect something working well here. not only cosmetics.


merci beaucoup de votre attention.






LBH

  • Hero Member
  • *****
  • Posts: 1.681
  • Karma: 52
I think perhaps there are two thngs to look out for. I think both create troubles.

1. It's clear that when using the new VC applications like Buchla and DX7, then the CPU rise with the number of voice settings, no matter you play even a single note or not. (EDIT - in a way it seems to be random especially in DX7, as my readings is'nt allways the same. - EDIT END) -The behavior seems to suggest something is wrong.
This behavior may also create audio issues like clicks and crackles, as the load mainly is put on a single CPU core.
This i would say Arturia need to have a look at. I hope they allready have an eye on it.

2. If AsioGguard is something like Studio Ones so called low latency monitoring system, then i don't think it work well with plug-ins that have more than a certain amount of CPU usage. Studio One recommend not to use it for instruments unless they don't use much CPU. Effect plugins that use more that 2-3% CPU (as far as i recall) can also automaticly shut down. You can say the system is not good for things that shall process in real time. I can understand why, as real time processing need to catch up with the stuff the system is affecting, and thus the CPU actually has to work a lot faster or at lower buffers. I never use it. A clean simple buffer setting that's the same for all things, is for me the right thing to use. Also because different buffers can cause sync issues in different situations.


BTW: I see your CPU speed is higher than the datasheet say the max speed on Turboboost is. Are you overclocking your CPU? If so it perhaps can create performance issues caused by heat unless you have a very effective cooling system.

I'm puzzled by the even load on all your CPU cores.  especially when the load is high for a single instance. You don't have more going on?
« Last Edit: May 30, 2018, 01:51:39 am by LBH »

Humanizer

  • Apprentice
  • *
  • Posts: 18
  • Karma: 0
hello again
thx LBH for your attention.

first, as i understood, ASIOGUARD is a mechanism that increase buffer size for tracks or vsti's that aren't in monitoring mode(no need to be played in realtime) it allows precalculate bigger chunk of audio before spit it out . thats why when monitoring is off, peak load is lowered. its a real benefit. when asioguard is disabled, peak load stay at the same level(higher).whenever track monitoring turned on or off.

Second, the processor is overclocked here. Cooling system is a custom EKWB loop. yes it's over the stock speed. But don't expect obtain fantastic result for achieve low latency with stock settings.

 Intel SPEEDSTEP is disabled and all cores are running at the same speed. that means that every core stay in sync at 4,7 ghz .that means there no switching of frequency clocks. its a very good thing because switching core frequency clocks can cause dropouts when buffers are very low.
 throttling can occurs only when alls core are at 100% when running a program like OCCT or a heavy duty CPU based rendering program like Blender. in this case throttling don't even exceed 1%. I project to delid and repast the proc soon to get rid of that.

In DAW context, i had running some tests: at least 250 tracks of reading/recording stereo channels , realtime processing.... bunch of VSTi's still at minimum latency, At 96kHz( lots of plugins sound drastically better in 96khz, like KORGS , NI's D16 and others) without blinking an eye. i experienced KONTAKT had achieving a least 4000 voices ....
no, We can't really blame here overclocking.


So..
 see on screenshot 3: you can read temperatures of the system: processor temperature here is ok at 42 Celsius degree for a cpu load at 42%. Ambient temp is at 25. Water at 29.
but look attentively: here we are running AND PLAYING 12 INSTANCES of MONOPHONIC BUCHLA with minimum latency,  in 96kHz ! it's not bad at all. considering the Bulcha is sounding very analogue, i suppose its calculations are very cpu consuming.
In this case, each instance CPU consumption of the buchla is distributed equally on each logical core used by NUENDO.
Its fine in this case

but look now on screenshot 5: only one instance in polyphonic mode (just only 4 voices) put all 14 cores used by nuendo abnormally loaded... causes sound grungelized( not really spikes and pops). on the task manager graph it seems that the CPU load is not divided equally between the logic cores but REPLICATED ! :o something definitively goes wrong here.

what is the conclusion?
need serious code optimization?. polyphony mechanism? need to do elegant programming, and maybe adapt to Market new mechanism (i9,Steinberg, Microsoft)?

bespite computers become always more powerful, developpers should keep an eye on performance impact of their products.
and keep in mind that all kind of AUDIO Programs should be expected to run at lowest latencies on Hi-end systems.

i don't have yet perform benchs on CMI V and DX7 but i expect trouble too. like the Matrix12 v.  :-\


 

LBH

  • Hero Member
  • *****
  • Posts: 1.681
  • Karma: 52
Hi Humanizer,

As i wrote in my last post, then i think Arturia should look at the polyphony behavior.
Then monitoring system in your DAW is a DAW issue i think. This Arturia can't do anything about.

This 2 thing seems to be the main issues in this case.

on the task manager graph it seems that the CPU load is not divided equally between the logic cores but REPLICATED !
This was what i refered to.


As a side note, then keep in mind, that there are differences between CPU's and the way to get the most out of them.

FYI: You might be interested in this article about overclocking for your CPU:
https://www.tomshardware.com/reviews/intel-core-i9-7900x-skylake-x,5092-10.html

Here are some lines from the article in the link:

"We did manage to achieve a stable 4.8 GHz overclock under the single- and multi-core Cinebench R15 benchmarks. However, our cooling solution was probably the decisive factor there. Realistically, 4.5 GHz should be achievable with an all-in-one liquid cooler."

"So, what’s the final verdict on overclocking Intel's Core i9-7900X? It depends on the quality of your cooler and its ability to move heat off the CPU as quickly as possible."

According to intels specs, then 4.5 GHz is the normal turboboost speed of your CPU:
https://ark.intel.com/products/123613/Intel-Core-i9-7900X-X-series-Processor-13_75M-Cache-up-to-4_30-GHz

No doubt about you are having a great system.


What do you exspect Arturia to do beside checking the polyphony and the generel CPU performance for their product? Sound like you exspect they solve everything including DAW issues and perhps others. I suggest you contact Steinberg about the monitoring issue.

What do you think the conclusion is yourself?

Humanizer

  • Apprentice
  • *
  • Posts: 18
  • Karma: 0
thanks for your interest Mr LBH. i appreciate your point of view.

thanks for your links  i had already read it few months ago. btw interesting to re read.

first i consider that the jump on the "real time peak" vumeter isn't an issue as we working with lowest buffers. real time peak and average load are different indicators. so for me there no issue concerning monitoring behavior of the daw. this jump is normal as i don't hear any audio problem when i switch it on.

my opinion is maybe the buchla ( and all the arturia products ) haven't been engineered to work well at 96kHz. as a majority of people use it in 44.1 and 48. at least not well tested. i find some of their products not very efficient ( like the convolution reverb of the gorgeous farfisa, compare CPU LOAD when it is on and off... and listen the glich)

Maybe the problem is related to how NUENDO/CUBASE handle plugins and processors. then if i got time ,i ll test on other daws like SONAR (now cakewalk i think?) or reaper, and in STANDALONE MODE to see the behavior. if i get no issues i' ll whining at steiny... :D

its a good idea to contact Steinberg but i bet they will say : "as you achieve good performances with other plugs , contact your plugin manufacturer." ::)

Anyway i think the fundamental problem is more theorical: digital signal processing on parallel processors. and its a big challenge today as cpu core count constantly increase.

and i expect Arturia (beside continuing produce marvelous plugins as they do) to overcome these math problems in concert with majors firm as Steinberg (who invent VST), Intel, AMD, Microsoft and Apple. in any case they are more able to do that than me ! i have done my part of the job ( i ve paid  ;D ).

...and less bullshit marketing à la "your experience your sound" please ::)

keep the good job !




LBH

  • Hero Member
  • *****
  • Posts: 1.681
  • Karma: 52
Hi Humanizer,

Thanks for your posts, inputs and observations.

On this page https://www.arturia.com/products/buchla-easel-v/overview you find the following note:
"Note that while our extreme modeling at the component level on the Buchla Easel V delivers insanely great analog sound, each voice of polyphony requires significant processing power to pull it off—approximately 20 percent of an i7 CPU per voice."

The problem as i see it is, that you have the high CPU usage no matter you play a note or not. The usage should as said depend on how many notes/ voices you play, no matter how many voices polyphony Buchla is set to.

As it is you have to turn off Buchla to save CPU. It's not enough not playing it, as it use the CPU anyway.


If needed then some workarounds in some situations can be:
1. If possible use automation to turn off Buchla, when it's not in use.
2. Do realtime Midi recording in a low samplerate and then render the track to audio using a higher samplerate if one want that, and then turn off or unload buchla to save CPU ressources. Perhaps you have other options.


I think it's important for developers of music production applications and CPU developers to solve the issues to use multiple CPU cores without getting to much latency. It's a long topic. But some things can be done, and it's about time doing something no matter how hard it is.
Also i certainly agree, that developers shall do their part of the testing. That said, then reports can help further.


I don't know if Arturias plug-ins use internal oversampling. A "Quality/ oversampling" setting on the applications could be useful.
A synth can for example run 96 KHZ or more even if the DAW project is 48 KHZ, depending on the used oversampling.
I would like Arturia to inform about this, and to implement a "Quality/ Oversampling" setting in their applications.


BTW: Perhaps it could be good to report to Arturia support directly - especially about the high CPU usage, even when no voices is played, as described.
I'm a user like you.

I still think the monitoring issue is about the DAW. Off course the issue is bigger in some cases, but it's still also about how that DAW feature work.


Best.

Humanizer

  • Apprentice
  • *
  • Posts: 18
  • Karma: 0
my workaround will be simple:

as the original buchla is monophonic, i ll set polyphonic preset to mono . thats all

if i need some polyphony: i will do with several instances rather than only one. and i ll  split midi voice to it.

thats ll be a bit cumbersome but after all , thats the way we do in the seventies !!

cheers

 

Carbonate design by Bloc
SMF 2.0.13 | SMF © 2016, Simple Machines