Ok, here we go:

Individual VST/AU/RTAS plugins will run on a single thread (= single CPU core) within the host. Many hosts use multiple threads in parallel to distribute plugins across the CPU cores.

Individual plugins however can start their own threads. E.g. disc streaming sample players can have one thread that's busy pre-loading samples while the dsp thread plays them. Or, a synth can have a set of threads waiting for voices to be played, and distribute single voices on various cores.

Multithreading has an overhead (e.g. context switches). This means, the total cpu load goes up as compared to single cpu. This overhead is very high on Northbridge CPUs and older - Diva might even perform worse on those, if multithreaded. It's very low since Nehalem and i-series Intel processors. That's where it rocks. We have no data on AMD though.

We'll observe this and refine it in time. But we simply had no luck at all with Core2Duos and early 8-core Mac Pros, no matter what buffer size, no matter what threading technology and no matter what data/code arrangement we used. Older machines (pre 2009?) are simply not up for low latency realtime multithreading.

Cheers,

Urs

@Urs: Somewhere you mentioned that certain transients will be different, whith a more analog character. Will this be implemented in v 1.1?

Maybe you could tell us a bit more about this, please.

Hi Urs ,

so wie ich es verstanden habe bringt das alles nur wirklich was ab Quad Prozessoren aufwärts.

Gibt es keine Möglichkeit das ganze auch auf Core 2 Duo Prozessoren zu Optimieren.

Meine alte Kiste ( 2,8 GHz Intel Core 2 Duo ) Imac läuft Super,das wäre schon schade wenn es da nicht viel

bringen wird.Multi ist wohl doch nicht gleich Multi?

mArIo

That's a looong story, best cut short:

Like every softsynth (and for good reason), DIVA puts her voices on standby when you can't hear them, i.e. if VCA level is down near zero. This saves a lot of cpu. Thus when you play a new note, that voice has to be woken up from standby quickly, and put back into processing. Even though it may sound ridiculous, this sudden wake up can cause a lot of artifacts which usually show in inconsistent transients.

Funnily enough, there are similar artifacts in analogue machines, so we haven't given that much consideration. There are however slight differences, and we think we've sussed out what they are. Basically, a Minimoog has that audible "thunk" attack only when you play big intervals, but not when you repeat the same note with Release off. And of course, it doesn't always happen. In Diva 1.0 it might also happen on retriggering the same note, which is wrong.

The main issue however is obvious: In an analogue synth, when you turn a knob or when you have lfo modulations going, these changes also take effect on the silent voices, because they are still "processing". In digital synths these changes take effect all of the sudden when a new note is pressed. Thus, if you switch a preset in a digital synth, you often get a couple of "thunk" transients until each voice has been processed to reflect the settings (unless the synth resets his voices anyway, which analogue emulations shouldn't)

You may think of it as a sudden change capacitor charges within the dc-blockers that a re distributed all over the place. Some synths don't really need them, then it's a lesser issue, but an analogue emulation is certainly full of these.

In Diva 1.1 I think we've done a great job to minimize this effect, while preserving an *ahem* low cpu footprint. In fact we're using tricks to "pretend" that voices are still run in the background, even if they're on standby. This improves the authenticity of transients by a great deal IMHO.

Urs

multithread works on Ableton Live, but not on Logic Pro here
i use a macbook i5 13" with lion 10.7.3

maybe an option to turn off unused oscillators would help reduce CPU hit ? i know about the different choices of oscillators, but that could still be useful i think

thanks
Last edited by flurp on Sat Apr 07, 2012 1:22 am; edited 1 time in total

Haben wir nicht gefunden, leider. Bei meinem 8-core bringt es auch nichts.

Bei einem i7 dual-core funktioniert es hingegen sehr gut.

Ob es an uns liegt, oder am Prozessor oder daran, wie die Systeme den Prozessor unterstützen, kann ich leider nicht sagen. Aber es scheint damit zu tun zu haben, dass sich die alten Prozessoren einen einzigen Datenbus teilen müssen.

The oscillators are not the culprits. They're pretty much as fast as Zebra's.

It's all about the filters, I'm afraid.

Thanks, Urs for your very clear description! Can't wait to try the official release of Diva 1.1. Keep up the good work!

sad but true

Hi Urs,

multithread also works well here, pops and clicks are gone (I'm using an older quad core Q6600). So congrats for this great work!

Unfortunately I also have some disadvantages using brainspawn forte 3.0 as host. Whereas everything worked fine here with Rev. 517 (the overload scenarios did only occour in cubase for me) I now have very strage effects concerning storage of DIVA parameters in a scene (forte can call up a bunch of plugs with there parametrization in a scene and there mapping to the MIDI-Controllers I use on stage).

When I store a scene with one or two DIVAs activated in it and call it up lateron DIVA randomly reacts with changed settings for some of the parameters: this can be a filter frequency, the tuning of the voices, ADSR Parameters or anything else. In some cases I even get hanging notes when changing form one scene to another even though no note on instructions were sent. To be sure I de- and reinstalled Rev 517 / 585 several times and always with the same results.

I confess forte is a very special host application and has a different behaviour than a DAW like cubase and I understand if this doesn't stand in the focus at the moment. But it would really be kind if you'd have a look on this when the main work for DIVA 1.1 is done as DIVA is also a great synthesizer for live.

regards, humphrey

Humphrey, this sounds like maybe the parameter list got mangled. We haven't checked that yet.

I'll put Forte on the list of host software for Fritz. He'll set up testing scenarios for all hosts that often come up in bug reports.

on an amd phenom II improvement is at least 30% (from a crackly 8-12 polyphony to a comfortable cristal clear 16 poly on my my most cpu intesive pad)

i do get a crash if i switch between accuracy modes and clickin on and off on multithreading while the notes are in play and with the cpu at the edge.
for now i just avoid touching those parameters when song is in play.

Thanks alot, really terific service at u-he!

Regards, humphrey

Have you considered adding another unision mode where all unison voices would share same filter? Also, adding paraphonic mode would be very very nice IMHO (not just for reasons of lower CPU load).

edit: what's the status with mod matrix (I would like 4 Tyrell-a-like mod slots even more than any other feature, and BTW if you can add midi CC as mod sources in mod martix, that would be even better).

Working very well for me in Logic pro (i7, MacMini). I can play one particular patch of mine that usually struggles at anything over 4 voices @ Divine (in fact I set it to 3 in my DIVA-Nation soundbank version) with 16 voices @ Divine now!