speeding up audio mixdown

Hi guys,

Most of our work in the studio are tv shows (30,45,65 minutes episodes) and audio mixdowns take a significant amount of time to process. (about 1/2 time of export lenght)
I’m wondering what could be a bottleneck here and how could I speed up this process.

On mixdown total CPU load shows 35-40% (i7 6850k)
network load is at 15Mb/s (we use synology 815+ for data storage) so very low

I wonder if changing the architecture to x299 i9 CPU could speed this up (more corese) or maybe I should go for faster single core speed CPU like i7 8700k?

what do you think?

greetings,
kamil

I’m guessing that if your CPU is far below maximum capacity, at 40%, then upgrading the CPU won’t help. I mean, it seems like the bottleneck is elsewhere. The things I would look at are;

  • Increasing the buffer size. I’ve seen it help export time.

  • Try moving one test project and render export from / to local SSD drives. If you see a difference then despite pulling only 15MB/s off of the network something is clearly different there. And it also begs the question of just what your projects look like; the amount of tracks and the amount of processing etc. I know that Intel’s platform architecture often runs several peripherals through the chipset, which might include ethernet, and that chipset is limited to PCIe x4 even though there may be far more lanes available (on Threadripper for example a lot of connectivity goes straight into the CPU). It’s nearly 4GB/s though so I guess that shouldn’t be a problem. Anyway, I would try local versus network and see if there’s a difference.

  • Go through your project’s plugins and see if there are particular ones that are ‘heavy’ one way or another. I know on my pathetically old system using multiple of iZotope’s RX Dialog Denoisers in realtime will bring the system on its knees, and on Pro Tools v12.x there at least was an issue with Equilibrium EQ which made it export slower than expected (during non-realtime exports). There’s also the off chance that a plugin “needs time” for one reason or another. I’m thinking of plugins like UA’s UAD plugins which I believe under certain circumstances slows down export speeds. I’m not sure if that’s because they’re already maxxing out their connection’s bandwidth or if they’re already at maximum computation speed, but that’s what I’ve heard. So if you suspect a plugin then maybe try to render a segment of the show with and without it.

Oh, and remember to report back and share your progress.

Thanks Mattias,
I will do a test following your suggestions.
will let you know!

greetings,
kamil

I did some checks.
Moving project to ssd drive from network drive speeds up audio mixdown but only few percent.
I work on a buffer 1024 and changing it to 2048 doesn’t seem to speed up the mixdown.
I think it’s the cpu. Overall CPU usage is about 35% percent, but only the first thread (CPU 0) of 12 is maxed out close to 100%…
Removing plugins, especially from izotope like mouth de-click, neutron2 or dialogue denoise speeds up the process significantly.

So the logical move would be to find a CPU that is fastest per core not to look for as many cores as possible - and according to benchmarks it’s 8700k.
Or maybe there is a way to optimise Nuendo and plugins to use all avaiable proccesing power from other threads and stick to 6850k I already have?

There is an explanation either here or on Gearslutz about how DAWs tend to load CPUs. I think it has to do with how plugins are allocated as well as the total load. So the only other thing that I can think of now is trying to “re-balance” the plugin load.

So if you for example have a signal chain that consists (partially) of 12 dialog tracks feeding one dialog group, and you’ve slapped many heavy plugins on that group channel, then I’d try to break them up onto different sequential groups, or move some processing back onto the original audio tracks. It’s a bit clunky and not intuitive, but I seem to recall that processes get allocated to cores partially according to where the plugins live. So that might be worth investigating.

(The above is what I’m thinking because you mentioned one core being close to 100%, which I guess could be a bottleneck.)

And actually on that note I’m wondering if there are any Nuendo settings that could change things. There’s the toggle of “multi-processing” (which I’m sure you’ve selected) and “VST guard” (which I know little about) etc under “device setup”… Have you messed around with those as well?

yes, I’ve been searching the websites and it seems that multiple instances of a plugin can be spread across the threads, but I’m trying to avoid the situation when I choose plugins for lower CPU rather then thinking about the sound.

VST guard helps in realtime processing on multicore systems and prevent from cpu spikes and audio dropouts and it works very well - I love this feature.

There is one interesting thing to investigate - intel turbo boost 3.0 - if I understand ths correctly - it speeds up the most used core for selected application - looks like a good solution, will report back after testing.

greetings,
kamil

Turbo can indeed be worth looking into. I’ve been looking at a new build and when reading up on AMD primarily, the interesting thing is just how things scale when doing a manual overclock and locking all cores to the same speed versus the automatic turbo…

Looking forward to hearing what you find out.

I realize this isn’t as much fun (not even close) as figuring out how to make one’s computer take fullest advantage of its processors… but for what it’s worth: In addition to music and post, I do a lot of audiobook production. The editing/post process for those puppies is audio janitorial at its most tedious, and there’s nothing worse than excessive, repetitive mouth clicks and glottal catches. So… iZotope RX and its plugin modules are a huge part of my toolkit (wtf won’t they bring back the Spectral Repair plugin? I’m still using v3 for that, but I digress…).

I used to run De-Click and De-Noise as realtime processes at export, but for a 10 to 20-hour audiobook it just got annoying. My “step one” these days for any potentially problematic dialog is to pre-process everything in the RX standalone app for basic, light-ish de-clicking, de-crackling [about the same level I’d be running in realtime, if I were], and sometimes even ambience matching. Obviously, I save all the originals just in case a glitch pops up, but to date I’ve never had to go back and fix anything because the RX processing messed up. They and the UAD plugins have been the biggest drag on non-realtime export (I generally spit it out with a 2048 buffer), and using mainly built-in Nuendo compression, expansion and de-ess (and in the case of triage, a Post Filter or two), the API vision and Precision Limiter, I usually get about 5x realtime from my exports. Saves a ton of time and aggravation in the long run, plus, makes the whole operation run much snappier, as those processor-intensive RX plugins can really impact realtime response, even at low buffers.

That’s a lot of yapping for a basic two cents, but there you have it…

Chewy

I do a lot of audiobooks as well. I agree that the Izotope plugins slow down the system. Sometimes stacking two on one channel slows it down beyond belief. Outside of fixing clicks manually while editing, I may run declick on my AIFF master mixes (along with denoise) before final conversions to MP3.