I just saw this thread from RADAR https://audiosex.pro/threads/gpu-audio-plugins-14-09-2022.66609/
I know that many of us have a powerful graphic cards, maybe because we work for music for movies, or we do video clips, or play games etc ..

Offload work from the CPU, using the same computer we already have maybe a good idea.

These guys at GPU Audio ( https://www.gpu.audio/ ) say:

What are your thoughts? Is this something you wish to generalize?

 

As soon as I'm able to use this tech with my Acustica library, I'm ALL for it.

Between GPU Audio and having AudioGridder, my current machine should last much longer than I expect.

 

First of all, I think it's a good thing to do something different or even better.

However, it also comes with additional costs. There are a lot of new plugins offered on the site, how much will
they cost and how good are their plugins compared to Fabfilter or other plugin manufacturers. This should first
be tested by independent musicians. Isn't it cheaper to buy a bigger, faster CPU? Why should I change?

CPUs like the Ryzen are very cheap and the expensive ones are actually graphics cards.
People with little budget use the onboard graphics chip.

 

Obviously this technology is aiming to people who already have a powerful graphic card, not to change anyone's computer configuration.

Also, about the price, I'm speaking about the technology, not about this developer particular offer. They are not the first who speak about implementing this.

Just imagine that DAW and plugins would be able to do this, let's imagine that plugin could take your GPU in account, if you have a powerful one, if not it would work normally as before.

 

We talked about this on another thread months ago. It would be really nice, of course, and it's already used in several other fields of CPU-hungry software.

The problem as far as I know is the latency (as in delay) introduced by the GPU processing which makes real-time work like DAW/plugins not possible with low latency in them. If you use a high latency there's already a couple of plugins that have implemented it. Like Nebula.

That being said I'd love to see some company pull it off.

 

Yes I can imagine that very well and I am happy about it too. I think in a few years, if you do your thing well, they will still be able to convince some and maybe in the distant future they will be able to have a bigger market share.

I think if you are a software professional, you will do something that you want to earn money with.
Let me give you a practical example. I have 9 Inserts FX plugins, I would have to change all my plugins that I use now. That's not going to happen, so I'm not the customer of tomorrow. So why should I break with the tried and tested and get to grips with new plugins.

 

The question to me is how useful is this going to be in reality? Most people already have CPUs that can handle a crazy number of instances of the example plugins they have running on GPUs (chorus, delay, EQ, convolver etc).

Almost all plugins that are slow today aren't slow because they run on CPUs, they're slow because the programmers didn't have the time or willingness to learn and make use of the vector instructions that all CPUs have had for the past 20 years. Those same programmers are not going to bother to vectorise their code for GPUs.

In my opinion, they came up with a technology before coming up with a use-case (for people making music anyway).

I think they're one of those proof-of-concept companies that exist to get acquired by much bigger companies that need to process insane amounts of data every second.

 

I think the use-case exists. The same people who use AudioGridder, Vienna Ensemble, Reaper ReaMote or Waves whatever-is-called.

But the technical challenges are big.

 

I think being able to partly outsource the CPU consumption to the GPU would be a good thing for mixing and could even be great for mastering where you probably can't make use of all cores/threads and are pulling out the good (aka CPU hungry) stuff. Also latency doesn't really matter here, so that wouldn't be an issue.

 

I think the latency is getting better on newer Nvidia cards, there are even low latency modes (in/out) finally. Needs the right display / outputs, so I don't know how practical this will be for audio, audiocards. But it seems more likely, @Xupito? Maybe there will have to be an (M1 style?) revolution to take advantage of parallel processing, big data / tables (famous LUTs for audio =)?

 

Last edited: Sep 17, 2022

NVIDIA GPU is used also for other purposes than graphics like Hashcat in Linux, so I guess in future for heavy vsts or physmod instruments developers will take for granted that additional calculation power.

 

I wouldn't guess this. CUDA has been utilized by linux penetration testers for well over 10 years, and we still see such slow adoption/development of GPU acceleration for audio. It's like they have almost no/little parallel progress in comparison. Relatively speaking.

 

I'm dreaming of a hybrid-solution, where all the calculating & heavy load stuff of our plugins gets outsourced to the cuda portion of an Nvidia card (like it does for transcoding videos) but latency-wise still the cpu corresponds to the audio card.

 

It's not only the latency but also the instruction set that's not exactly made for audio DSP.
Offloading tasks that can live with such limitatons (lengthy ffts/convolutions etc) can well be processed on a gpu (even the Raspberry Pi can do it surprisingly efficiently!) but as a general DSP booster it's not really sexy for developers, not the least because you can't just split DSP calculations into several threads if a lot of processing has to take place in a fixed order.

 

Without going into too much details, we either need plugin devs to impliment this, or daw devs to do so, doing all this work just for their own plugins isn't smart, so for now i hope this get's the attention it needs so major devs jump in, especially with the current gpu prices it's kind of the perfect time for that...

 

You're right, audio coders are reluctant to use other that CPU, and yet the GPU option seems to me not that far fetched, maybe in future when physmod will step into a new generation, and most people can't afford a Frontier or Fugaku.

 

I agree with you. This could only become a great thing if it's not specific to DAW or plugin brand, but if it became part of the general code for all plugins.
If they could integrated this into vst3 or CLAP for example.

 

Well said. This is one of the main issues if not the biggest.
GPUs process data in a massive parallel way. A medium-end GPU has more than a thousand of CU (compute units). And the CU is effectively acting like a CPU core. It hurts only by thinking how do you code several DSP stuff.

 

Last edited: Sep 18, 2022

Same than audio DSP : CPU are now so powerful it is LESS and less interesting to use anything else than computer CPU.
See Quad Cortex / Neural DSP plugins: they use THE SAME technology, in different code (of course).

NDSP plugins use only some % of CPU for each instance. So what would be the point for them to create an external DSP or GPU card ?
See how many ppl dropped external DSP audio card solutions last years.

It was a good thing decades ago. But now it is overkill, especially devs side.
Apart Nebula users with an hundred instances for a complete console simulation, i don't see the point.

As a side note, we can run plugins on external computer "boxes" for decades.
We can buy a second hand PC and run extra power on it with Vienna Ensemble Pro, Audiogridder ...
And sometimes, it cost LESS than a GPU (because GPU market is collapsing for 2 years now).

GPU for audio was a 2000's idea ... and it failed in the real world.
There is no way it will come back.
Same with audio DSP.

As a side note, i'm selling my LAST ever hardware guitar sim (with DSP) ... and will run ALL on desktop and laptop computers.
Flexibility, price ... everything is better.