Hi Guys,
What do you think about this?:

https://www.gpu.audio/

 

It's a cool idea, but i have my doubts. Afaik this thing only works with their own plugins, which isn't bad, but they don't really seem to be replacing anything CPU intensive in practical terms, meaning you should be able to load a lot of this type of plugins on the CPU without any issue.

If they do offer an advantage besides CPU consumption, as in if they sound better than a CPU intensive plugin you have to the point you can actually replace it, then that would be awesome. It's hard to tell though, since i couldn't find audio demos (besides the reverb) or any info regarding what's really in the bundle, besides some loose images here and there, and apparently my graphics card is not compatible so i can't test it myself.

I'm aware it's a new thing that's still under development, but i don't see it being a gamechanger plugin-wise unless they really make some top-notch plugins that actually bring some benefit by being run outside the CPU which i guess they will eventually, but considering how even UAD is slowly making it's way to native, will it really be necessary by that time?. Still, the technollogy itself and the idea of using GPU resources for audio is definetively something truly awesome, and i hope it keeps moving forward and doesn't fall into obscurity.

TL;DR: Idk, the idea is great, but to me there isn't a lot of substantial proof that it will actually bring any benefit to the plugin world as for now, so it's too soon to say if it's really useful or not. That said, i don't know a whole lot about the project so maybe i'm full of sh*t.

 

Can they (will they?) make a wrapper or sth, or is that impossible? Maybe we can just hope they license the tech like game physics things? Anyway, exciting af; just been waiting for sth like this to be adopted by DAW devs, have no idea why they're still (just) using cpus, tbh.

 

New Company, GPU Audio, Wants To Turn Your Graphic Card Into A Powerful Audio DSP - MARCH 22, 2022

GPU Audio is a new technology company that says that they have created the world’s first novel GPU-based, full technology stack for audio processing.

They say that they want to let you use the parallel processing power of graphics processing units, aka GPU cards, found in modern laptop and desktop computers, and let you use it for audio DSP (digital signal processing).

“Our mission is to make GPU Audio the next standard of audio processing, so that music and audio production can stand up to the demands of 21st-century content,” said co-founder Alexander Talashov. “GPU Audio holds the key to fast, easy, and unlimited power needed to allow audio producers and adjacent industries to participate fully in the future of content, production workflows, audio tools, software engines, and more.”

Features of GPU Audio:

• Low-latency VST3 performance regardless of channel-count
• Real-Time (instant) Audio processing
• Performance gains for AI and ML algorithmic use cases
• DSP power that is orders of magnitude greater than CPU
• The GPU Audio ecosystem is composed of both consumer and business facing models of engagement.

For the audio producer, they plan to offer a proprietary suite of VST3G plugins, serving up a full range of standard music production tools, including spatial audio tools, developed in collaboration with Mach1.tech.

For the developer, they will offer a fully-built, modular SDK, allowing them to create custom DSP products and implementation.

As an initial step, GPU Audio has released an Early Access plugin, alongside their keynote at NVIDIA’s GTC conference in March 2022, which inaugurates standardized GPU powered VST3s for the music and audio production community.

The Early Access community is focused on benchmarking and bug-squashing, in preparation for the beta-suite release, in early summer of 2022. This plugin demonstrates proof-of-concept of GPU Audio on one of the most demanding of algorithms: FIR convolution reverb. By offloading DSP onto a computer’s local or remote GPU, it dramatically increases performance, by allowing real-time parallel audio processing, across hundreds of channels and VST3 instances, without added latency.

You can check out the Early Access Plugin at the GPU Audio site.

 

being it is swiss made my guess is that they gonna charge some crazy-high amount of cash of the same old la2a/1176/ssl/neve/fx emus. also, it's one thing that their plugins will utilise the gpu in a way but what about the daws? how will they know to share the code? daw makers won't rewrite the audio software engine just to please this company.

 

The idea of outsourcing audio processes to GPUs is as old as the hills. If I'm not mistaken, there was or is this possibility in nebula, reaper has a corresponding function for nvidea GPUs, etc..
Let's see if this catches on with other developers throughout the industry. So far, the implementation has always been very "experimental".

 

Not that relevant anymore IMHO now that we have M1/M2 but might be interesting if you really need to use very long impulse responses.

Like @Sinus Well said, it's an old idea that mostly suffers from the GPUs being very high performance but not optimized for audio - except a few applications like convolution or FFT.

(BTW, even a Raspberry Pi can have surprising FFT/Convolution performance when it's processed on its GPU. Some latency involved though.)

 

Last edited: Jun 9, 2022

that company name already doesn't sound trustworthy

on more serious note, GPU by design does not have direct access to system resources like RAM or SSD, so latency overhead will always be too high, compared to all-in-one SoC like Apple Silicon

 

I'm in beta testing for their first plugin and it works surprisingly well for such a young technology.

Still ways to go but it will be great if we can utilize that fancy ass expensive GPU for music, instead of it just staying idle.

The trick is to correctly use a GPU's parallel processing for audio, which is highly dependent on serial processes (think FX chains etc). We'll see.

 

I already saw this one on FB.

The real surprise is that nobody thought about doing this before.

I already thought about this a few years ago when I saw all this Bitcoin mining craziness that was sending the GPU prices straight to the roof.

Now, if this company succeed (and I don't see any reason why not), it won't take long before UAD go bankrupt.

 

Oh yes, various devs have tried it with varied results during the years. This was actually a bigger thing 10-15 years ago than it is now because we have far better CPUs now. Well, and GPUs, but GPUs were always far more powerful than CPUs, that is for parallel processing [which plugins don't use], so plugins have to be written differently to be able to process audio serially with low latency. Big processing latency is still the biggest problem with plugins using GPU.

 

Last edited: Jun 11, 2022

I mean... WHY though? A GPU will cost you more than a high-powered CPU. And CPUs get you 200-300 + tracks these days, what will a GPU give u that will make u say, "YUP I NEED THAT". Not to mention like 90% of u monkeys still be using Intel 4700 CPUs and run windows 7, cause aint nothing worse for technology like... more technology. This tech might've been decent 10 years ago (when development likely began), but I'm pretty sure its time has came and went.

 

A new idea? lol.. yeah right.... anybody remember Nebula Correll /server setup?

 

Regarding doing DSP on GPUs, there's two big problems : Parallel processing and unpredictability.

In a nutshell, a CPU has way fewer cores than a GPU, but they run much faster and at a lower latency. Because DSP is all about running discrete signals through equations, you need to process a discrete signal one sample at a time, you can't really split the calculation easily and without a TON of workarounds.

Also, DSP needs to be 100% predictable, we all need our DAWs and plugins to have sample accurate processing. GPUs, because they have so many cores, can be unpredictable due to a high risk of interferences between cores. Interferences can occur spatially (due to shared memory access), to temporally (due to jitter). Controlling that necessitates spatio-temporal GPU management, and that means adding a lot of latency. However there are ways to do that quickly (here's a paper on that : https://escholarship.org/uc/item/0qp3s44h), and that probably is what GPU Audio is all about.

However there's not a lot of usecases in which a GPU is really useful, because there's not a lot of algorithms in DSP in which you can split the processing on hundreds of cores. The main ideas that come to my mind are plugins that feature AI learning (not all AI based plugins do learning) and rendering IRs but that's about it. @Sinus Well said that Acustica tried this on Nebula, and I can understand why for that specific technology.

 

Last edited: Jun 10, 2022

i have enjoyed reading this, thank you.

 

It's no surprise, because it's not true that no one has thought about this before. Devs were looking into doing audio/plugin processing on the GPU back when laptops started to get discrete GPUs as a potential source of extra power. I remember reading some devs feedback on this probably around 2010-ish. The devs that looked at it just found it wasn't really viable at the time, due to various technical constraints, and the gains didn't seem significant enough to bother with and it never went anywhere.

However, tools and frameworks are better these days, so we'll see how this one shakes out this time around, as so far, we don't know what the possibilities and what the limitations are of their novel new technology.

 

Last edited: Jun 10, 2022

As several said: latency. For real-time stuff that's a deal-breaker. For many other uses the GPU has been used since quite a few years ago. Mainly Nvidia because they have a solid software SDK.

1. We all know it's already important for CPUs with several - but mainly sequential - cores. The famous one-core speed. Now imagine that with hundreds of compute units (shaders, compute units, cuda cores, whatever)
2. Add to that two copies of memory between RAM and VRAM many times per second as dictated basically by the sound interface/DAW latency.

Also, VRAM has always more latency than RAM while more speed. That's why is always one generation or two ahead of CPU RAM (ddr6 vs ddr4 these days)

So it's not worth it for real-time stuff. Well, if you set the latency to 4096 samples we could talk... or only for offline mixdown render (too much trouble). Also for Melodyne most intensive tasks like the initial analysis (again too much trouble but that would be dope).

 

Last edited: Jun 10, 2022

Amm.. this "revolution" is dragging on forever.. are these guys the braingines? we already had the discussion

https://audiosex.pro/threads/braingines-use-your-gpu-as-a-dsp.57628/

..so, it didn't really cause any frenzy irl, did it? ;)

Thanx for all the insight above though.

 

Anything which can consistently improve audio latency is a good thing for all musicians and engineers.
HOWEVER - until manufacturing improves with part availability, at present, GPU's are way overpriced since COVID.
The GPU market in price needs to settle back down for this to be realistic at a bulk user-friendly level. High end GPU's that sold new for around $1,000 two years ago, since COVID, the current new equivalents are 2-3 times that in price.

 

again? thats a late 2002 concept. often tried. more often made to fail by audio proprietory hardware market bullies (avids hdx, uads ancient arm blackmail)... the market is too small and the kickoff with "custom" boring plugins already proves that they couldnt get big players on board so far