GPU Driven | The Future Of DAW Processing Or A Dream ?

[logifuzz-vst-plugins]

Beyond audio and the usual graphics/video/crypto mining, are anyone putting their GPU cores or extra memory to some other use?

They were utilised in 2004 for the Spirit Rover on Mars and NASA's GPU cluster at the NASA Center for Climate Simulation (NCCS)

https://www.geek.com/games/nvidia-techn ... rs-555177/
https://sciences.gsfc.nasa.gov/600/high ... /gpus.html

Looks like the first application its just our traditional graphics usage, the second one is a very interesting project for parallel computing!

[telecode]

Beyond audio and the usual graphics/video/crypto mining, are anyone putting their GPU cores or extra memory to some other use?

At my work we have a POC for trying to use GPU computing for statistical data analysis. Basically, some of the jobs take 2 or 3 days to complete, so we are trying to see if there is a way to use GPU computing to see if it can reduce it to less time to complete. It hasn't really taken off yet due to too much other stuff that needs to get done and not enough human resources available to devote a person to try to get it to work properly.

Interesting project, mind me asking what programming language you guys are using to do the GPU parallelism? This would be a proper academic project if the statistical task can be benefited from multi process.

I don't personally program it. They use R and Matlab. I just work on the infrastructure.

http://www.r-tutor.com/gpu-computing
https://www.mathworks.com/solutions/gpu-computing.html

Personally, I really don't know where GPU computing will go. Its all stuff that is being hyped by companies that what to sell hardware, but the days of buying big honking expensive hardware and then dealing with throwing it away in recycling after 7 years are coming to an end with the likes of AWS and Google cloud solutions.

[tenshin111]

GPU processing isn't that different from what UAD is doing with their hardware and plugins, so I can see it work.

It is different. UAD cards use dedicated dsp chips for audio processing.

[THE INTRANCER]

Around 10 years ago, Nvidia posted this PDF explaining it... it's quite colourful.

https://www.nvidia.com/content/GTC/docu ... _GTC09.pdf

[JerGoertz]

Around 10 years ago, Nvidia posted this PDF explaining it... it's quite colourful.

https://www.nvidia.com/content/GTC/docu ... _GTC09.pdf

Yeah, whatever happened to CUDA? Seemed like a huge deal with a lot of potential (and Nebula was using it IIRC) and now, nothing.

Edit: After reading the thread, I'm guessing it had to do with latency.

[PeterP_swe]

GPU processing isn't that different from what UAD is doing with their hardware and plugins, so I can see it work.

It is different. UAD cards use dedicated dsp chips for audio processing.

In what way is it different?

There's nothing on their website* about specific audio extensions and I can't think of what special asm instructions you'd add just for audio. I'm genuinely interested if you have more information and I would love to read about it.

*) https://www.analog.com/en/products/proc ... sharc.html#

[Xenakios]

In what way is it different?

UAD can design their whole hardware specifically for audio needs. The DSP chips themselves don't necessarily have to be audio specific, it suffices that they have good enough floating point and memory access performance. GPUs are designed foremost for doing video/graphics that will be outputted into a display device directly connected to the GPU. They are not great for doing heavy 2-way I/O between the GPU and the host system in real time, which audio will typically need.

[PeterP_swe]

In what way is it different?

UAD can design their whole hardware specifically for audio needs. The DSP chips themselves don't necessarily have to be audio specific, it suffices that they have good enough floating point and memory access performance. GPUs are designed foremost for doing video/graphics that will be outputted into a display device directly connected to the GPU. They are not great for doing heavy 2-way I/O between the GPU and the host system in real time, which audio will typically need.

Sure, every time you want data back from the GPU we're talking about 10-20 ms * in my experience, which is significant for real-time usage, but if you're in the mixing/mastering stage then it's usable.

And it's not per plugin, so you could in theory max out the GPU with a ton of plugins in the same batch without increasing it.

GPUs have been used to do physics, encoding and all sorts of calculations for a while now. I definitely think it's useful outside of graphics.

*) Those numbers are a few years old and may not have been optimal then, but I don't expect that to have changed drastically.

[blobbs]

It is possible with OpenCL. There are working examples on github. But now Apple deprecated OpenCL, because they suck. It was not old at all.

OpenCL is basically deprecated across the board in favor of more modern graphics apis that include compute shaders - so in the case of Apple you would use Metal compute shaders, or on linux or windows you could use Vulkan instead of OpenCL. So it's not like GPGPU is going away, instead it's been rolled into unified GPU apis (other than CUDA I suppose)

[blobbs]

In what way is it different?

UAD can design their whole hardware specifically for audio needs. The DSP chips themselves don't necessarily have to be audio specific, it suffices that they have good enough floating point and memory access performance. GPUs are designed foremost for doing video/graphics that will be outputted into a display device directly connected to the GPU. They are not great for doing heavy 2-way I/O between the GPU and the host system in real time, which audio will typically need.

Sure, every time you want data back from the GPU we're talking about 10-20 ms * in my experience, which is significant for real-time usage, but if you're in the mixing/mastering stage then it's usable.

And it's not per plugin, so you could in theory max out the GPU with a ton of plugins in the same batch without increasing it.

GPUs have been used to do physics, encoding and all sorts of calculations for a while now. I definitely think it's useful outside of graphics.

*) Those numbers are a few years old and may not have been optimal then, but I don't expect that to have changed drastically.

in my testing with images with new hardware (RTX 2070) I am seeing upload and download times of 1080p YUV 4:2:0 buffers (so roughly 3MB) around 0.7-0.8ms ... looking at the numbers actually I may have something weird going on in my measurement so I will come back and edit this post in a bit after i investigate

[Ah_Dziz]

So 1 ms latency at 96k. That's 2 ms at 48k. I wouldn't find that acceptable

Sounds OK to me - Apparently that's the same time it takes from a drummer hitting their snare to the sound hitting their eardrums

Not OK if you want to monitor with FX while recording live.

I’ve found that up to 256 samples is damn near unnoticeable for most people through their headphones at 48k. I’ve tested lots of different setups (just to see how far it can be pushed before anyone says anything or I notice a change in the performance) through the years with folks in and out of my studio. I can’t deal with much more than 128 for playing any keys that need accuracy myself, but i know it’s there. If you don’t tell people they don’t tend to notice much. I would be amazed if a 2 ms difference would be anything to anyone used to digital equipment if you didn’t mention it.

[v1o]

Are there any proof of concept VSTs or DAWs that run on the GPU?

[cthonophonic]

Liquidsonics made a couple free ones using CUDA, but it doesn't look like that made its way into any of their commercial products:
https://www.liquidsonics.com/software/reverberate-le/
https://www.liquidsonics.com/software/filtrate-le/

[planetearth]

Liquidsonics made a couple free ones using CUDA, but it doesn't look like that made its way into any of their commercial products:
https://www.liquidsonics.com/software/reverberate-le/
https://www.liquidsonics.com/software/filtrate-le/

I was just about to mention them when I saw your post. I've been following this topic for about 6 years now, and other than their free stuff, LiquidSonics doesn't seem to have used CUDA anywhere else. (And neither has anyone else, from what I've seen.) Reverberate LE has a built-in latency of at least 512 samples, so a few more milliseconds of latency isn't really going to matter much.

According to their website, the zero latency CPU edition can be delayed up to 8192 samples, but otherwise offers "zero" latency. But the NVIDIA CUDA GPU edition can't do less than 512 samples of delay (up to 8192:

"Zero latency CPU edition (or up to 8192 samples for lower CPU usage)
Low latency GPU edition (512 – 8192 samples delay)"

I guess they determined that implementing CUDA processing in their commercial products wasn't worth the performance hit. They didn't even upgrade the CUDA-based, "LE" versions of their plug-ins to 64-bit, which is another sign that they don't think there's a future in CUDA processing for plug-ins.

It's an interesting concept that keeps raring its head on forums every so often. But from everything I've read about it, it will never be useful for plug-in processing or anything that's time-sensitive, apparently. While some may moan about it, I'm just glad I don't have to worry about getting a powerful graphics card now, just to run more plug-ins.

Steve

[EvilDragon]

So 1 ms latency at 96k. That's 2 ms at 48k. I wouldn't find that acceptable

Sounds OK to me - Apparently that's the same time it takes from a drummer hitting their snare to the sound hitting their eardrums

Not OK if you want to monitor with FX while recording live.

I’ve found that up to 256 samples is damn near unnoticeable for most people through their headphones at 48k. I’ve tested lots of different setups (just to see how far it can be pushed before anyone says anything or I notice a change in the performance) through the years with folks in and out of my studio. I can’t deal with much more than 128 for playing any keys that need accuracy myself, but i know it’s there. If you don’t tell people they don’t tend to notice much. I would be amazed if a 2 ms difference would be anything to anyone used to digital equipment if you didn’t mention it.

It's not enough when multitracking live vocals or guitars, for example. You need the tightest possible latency then.