CPU Performance vs. Real-Time Performance in Digital Audio Workstations

[benny van de locht]

I am about to start a series of tests: One of them is comparing the performance of the external GPU with the internal. It is on a MAC but it will give an indication

regards

Thanks for the information Frank, great that you contribute so much to the comunity!

[Jostein]

I found a nice video on the topic from Molten Audio here:
https://www.youtube.com/watch?v=ojnnP_GXNaM

That's a nice one!

---

The tool from Fons measures the audio (roundtrip) latency and is basically built into Ardour/Mixbus for calibration in the audio setup dialog as well as for measuring the latency of channel inserts.

Cool, had no idea that it was Fons' stuff that was build in! :-)

I did not think about that it's round trip only, but on the other hand: That's the only thing I care about. Thankfully, I never had to go deeper in order to fix latency, the usual optimizations and careful choice of HW have been more than enough. Even my modest laptop is so far fast enough for recording.

[johne53]

Rather than replying in the Windows forum I decided to reply to this in General Discussions (where it might also attract the attention of some Linux users)

Does the soundcard and driver really have any effect on the computers performance when it comes to plugins? Reaper is the only other DAW that I have used on my current setup and it runs with practically unlimited plugins. I did a session with 8-12 softsynths and BFD sampled drums, I threw several compressors, limiters, reverb and EQ on all tracks and I was up to about 15% CPU usage (in reaper)

If any of you have used both Mixbus and Ardour you'll be aware of how much higher the DSP readings are with Mixbus. But of course, Mixbus is offering EQ and compression on every channel. For example, I have a session here with 26 tracks. If I load it into Ardour (on Windows) I see a DSP reading of around 3% (that's when using the Jack backend and a buffer size of 2048). The same session loaded into Mixbus (with the same backend and buffer size) gives me around 18%. So one day I wondered what would happen if I loaded the session into Ardour - and then I placed an EQ plugin and a compressor plugin on every track. I chose the EQ2B plugin from OvertoneDSP (basically, because I've always considered it to be an efficient plugin).

With EQ2B loaded onto every track, Ardour's DSP figure jumped to a whopping 32%. So then I tried the same experiment - but this time using Harrison's XT-EQ plugin (LV2). That gave me a slightly lower figure of 29%. I didn't even bother adding compression as I'm sure the reading would have been well above 40%. My machine's only a lowly dual-core Intel but I'd be astonished if I could run any DAW with hundreds of plugins. It's obvious that the plugins themselves are eating up significant amounts of CPU (either that - or there's something very wrong in the way that Ardour's interfacing with them).

What would be really, really, really interesting would be if someone here could repeat that experiment on Linux and let us have the results.

And if you're one of the lucky ones who can run hundreds of plugins, what CPU / video card / sound card are you using?

[johne53]

As a further experiment (see my previous post) I just installed 8GadgetPack from Helmut Buhler. It includes a very handy CPU monitor.

On my system, adding any EQ plugin adds around 1.4% to my CPU usage (not the Ardour DSP reading - the actual CPU usage). So 26 x EQ plugins will add around 36% to my CPU usage. It's about the same for either OvertoneDSP's EQ2B or for Harrison's XT-EQ.

Ardour (just on its own) runs my CPU at about 7%. So a 26-track session with EQ on every track runs my CPU at about 43%. This is with an Intel Pentium G620 (dual core) running at 2.6GHz.

Admittedly it's not exactly a beefy setup but it's clear that the maximum number of plugins I could run would be in the region of 66 (at which point, the plugins themselves would be consuming about 92% CPU).

If nothing else, it tells me I need a beefier system!

[bapu]

So, no one here has ever created a "benchmark" project with tons of tracks and tons of plugs (even if just one of the essentials) and offered it for download?

Funny that.

[johne53]

A benchmark project isn't really feasible because of differences in the way multithreading works between different OS'S and different CPU's. For example on Linux it's quite simple to allocate threads to a particular CPU core. But although that's (technically) feasible on Windows, it's much more difficult. Windows prefers to make those decisions for itself.

The problem lies in the way different manufacturers handle requests to get the current time. Intel processors use a single timer which operates across all cores. But AMD processors use different timers for each core. If the OS is switching threads between different cores (as Windows does) it's possible to request a time - then (a few milliseconds later) to request the time again and get a time back that's hugely later (or even earlier) than your previous request!

Any kind of benchmarking that relies on knowing the time will therefore give erroneous results with some CPU's and some OS's. The measured performance might seem wildly different - even though their observable performance might not be very different at all.

[benny van de locht]

I just installed the Marian Seraph 8+ WMX.
It's a highend pci-e card with decent drivers that made my dsp percentage on the same project decrease from 93 to 60-64 % using the same buffersize and samplerate.
I tested it against an Akai eie pro usb device, I set the playback buffersize in 32C to max with the Akai, and at the "large projects" selection with the Seraph... so the result can be even slightly better.
To me that is an obvious improvement of realtime performance.
Using task manager, all 8 cores of my i7 4770 are getting an average usage of about 65%, which used to be less before.
The project counts 50 tracks with another 90 third party plugins. Some cpu hogs, most of them normal in demand.

[Ben@Harrison]

I think a reference session is a good idea. It should use all Harrison plugins so that the same session can be run on anyone's computer: Mac Windows or Linux.

Then we can make some comparisons between different computers, OS's, and I/O devices.

[Frank]

I think a reference session is a good idea. It should use all Harrison plugins so that the same session can be run on anyone's computer: Mac Windows or Linux.

Then we can make some comparisons between different computers, OS's, and I/O devices.

I second that !

Can you define one so we can make comparisons ?


regards

[Ben@Harrison]

I can't this week; we are working towards the v3.4 release. But I think anyone could do it.