Massive Parallel Computing for Digital Audio Creation?

NVIDIA is now offering a personal supercomputer utilizing GPUs in a desktop format. The GPU card itself is several grand, and requires that a unique subset of C++ be written to utilize it (CUDA), but can apparently smoke the toast out of massively repetitive computations. While these cards were originally intended for graphics processing, CUDA opens them up to be supplementary processors.

This leads me to my burning question/idea:

Are software synthesizer developers planning on making use of this processing power? Has anyone heard of any work being done on the subject? The opportunity of running unlimited software instruments and effects (to the tune of several thousand, if such a thing was ever desired) or new computationally intensive instrument modeling or effects never before possible is very inviting.

Interesting...

They are reporting "up to 250 times faster than standard PCs", "960 cores", and "4 TeraFlops", and it comes with 16GB of dedicated memory.

Are you sure we're talking about the same thing? I believe this is on quite a different level than the SHARC 21369.

You might check out personal supercomputing for what I'm talking about.

Also, where are the KVR forums? What does KVR stand for?

Cello,
yes well aware of the product its been out for a yr now or longer. you are buying into the marketing hype.

again its nothing more than either an AMD system or an intel system with multiple video cards.
and at the present completely useless for audio and mostly video.
while alot of video stuff like Adobe CS4 has added Cuda support we are a long way from it being realistic to consider GPU processing which has been around for a dozen yrs now.

its more designed for number crunching than use for media (past gaming)
both AMD/ATI and nVidia have add in cards that can do this as well in fact ATI thats all they offer, while nVidia is selling complete systems.

do a search on kvr for cuda.

Ahh.
If wikipedia is correct. OpenCL is at the bottom of all of them.
That's good news.

http://en.wikipedia.org/wiki/OpenCL

- zevo

I'm wondering if this is already a technology behind the curve, though. I have 8 cores in my current machine. My next machine (3-4 years from now, hopefully) I suspect will have 16 or 32 cores.

zevo,

These cards are running 240 cores all at 1.3 GHz. The chips are simpler, but I believe that still puts them about 5 doubles ahead. So, assuming everybody keeps up their respective exponential growths, when CPUs hit 16 cores, GPUs should be hitting 480 cores, 32 cores for CPUs, and GPUs come in at 960 per chip.

yowzies.

In the past, GPU instruction architectures have been optimized for graphics operations: triangle/polygon definitions, coordinate transformations, geometric projections, colorspace mappings, and textures. Not all of those operations translated well into general-purpose computing. To make them work with any useful efficiency, the language compiler had to be "tweaked", and the writer had to know the code constructs that yielded the highest throughput. Often, what was needed to get done for the desired signal processing task would not fit that tortured computing model. The result is inefficiency, which translates into "not-all-that-the-theoretical-hype-would-make-you-think-is-possible."

That said, perhaps the newer-generation GPU/chipsets are either more flexible, or more powerful, or both. Another possibility is that the new compilers can better translate language constructs into GPU primitives. As with all technology, anything's possible, but the devil's in the details.

Hi all,

Count me amongst the skeptics. Some concerns upon reading about this a couple of months ago:

1) As others pointed out, only certain types of computations are amenable to parallelization. To be sure, some of these are important, but it isn't as though everything could be re-routed to the GPU.

2) Normally the GPU is a write-only type of situation as far as computation goes: Here are some data, do the computation and draw something on the screen. Don't call me; I'll call you. With audio, you need to get the results back out. It seems to me at first glance that you might need a CPU or two just to handle the results.

3) The software development kits (SDK's) are aimed at specific hardware, or so it appears to me and almost necessarily is the case. That isn't good. The first thing I would do is try to abstract all the stuff in any one SDK to protect myself.

4) An important point for me, but probably not for others: Single-precision computations. Speaking generally, you cannot do serious numerical work with single precision. Even doubles can cause problems. Admittedly with audio, this if often OK (e.g. record, mix, "master").


What I would like to see INSTEAD is an audio PU or APU. This device could be given world coordinates, boundaries, materials data, results from some Newtonian physics problem, then compute the sound from (say) a marble column collapsing as the result of being hit by a pickup truck travelling 90 mph being driven by a drunk farmer. This would include everything from the crunch of the pickup truck to the scattering of tiny pellets of marble.

Now that would get my attention.

Regards,
Dave Clark

Roflcopter

...even if I model the properties correctly, the result will be largely random.

<snip>

But it would not be better than existant top-class sound effects of today.

Hi Roflcopter,

I think you are misunderstanding the advantage of having this type of processor, and that's partially my fault.

I don't mean to create the sound effect of *some* truck crashing into *some* column and all of the resulting sounds, but a *particular* situation, a particular problem such as may be presented in a video game in real time modelling when keys are hit in a particular manner to cause a very specific collision and its consequences to be modelled in detail. Does glass break? When? Do the tires go flat? How much noise does that make? Is there any skid? What kind of skid noise would there be? Is it raining? Do parts go flying? If so, where do they land and how much noise does that make?

A GPU does the same thing for optics for a particular view of a situation, not just some generic view of that situation, so I thought that the intent was clear here. Sorry if it wasn't. With this type of processor, you could slightly change the angle of approach and probably get some completely different audio results, just as you can get completely different video results with a GPU with small changes. That's the whole idea, actually. Small changes can have large consequences on subsequent game play, the "butterfly effect" you mentioned, so you need to model them correctly. You don't necessarily want to accept some generic collision, although that might be fine for a sound effect in some song.

Regards,
Dave Clark

In sound production for movies I think the fakes are probably considered more effective in blind tests, however inaccurate they may be...

Hi Roflcopter,

I would not be surprised by this. The average listener's sense of hearing is --- well --- not overly developed so indeed they don't hear the difference. In Edison's day many folks claimed that they couldn't tell the difference between early wax cylinders, foil, or whatever was used for comparison, and the real thing. Perhaps Edison wasn't entirely forthcoming about these comparisons, but such results are not too terribly surprising. Personally I hate the current sound systems used in theatres. They are unbelievably unrealistic. Even stupidly unrealistic.

My interest in an APU is that I just want to offload all my acoustical and physical modelling stuff onto someone else!

Then, just like everybody else, I can simply draw a room, place objects in it, put physically modelled sound sources in there, insert a listener, feed the instruments some MIDI or other score, and presto! Just like the real thing, this time "Fer real, fer real." This stuff is by no means limited to virtual reality scenarios.

On computation: There are lessons to be learned from the history of graphics where very pessimistic predictions were blown away by enterprising programmers. Something like Doom wasn't supposed to be possible on a mere PC of that time. Radiosity lighting was in the far, distant future. And so on.... Yes, there is a lot of computation, but there are techniques for cutting this all down by orders of magnitude. In my own experience, using Green's functions is a good example whereby one obtains a solution for one particular location for each of source and user (plus boundary and initial conditions) rather than using finite difference or finite element methods to solve for everything everywhere. For a concert hall, such an approach (albeit limited to geometries where the Green's function is known or easily calculated) the size of the problem is reduced from a "challenge problem" (US Gov lab term for "damn near impossible" for any existing architecture) to something you can do on a P4.

Now to return to the normal station....

Best regards,
Dave Clark

Meaning: I think I can make an efficient wavetable generator that would allow one to add dynamic patterns of harmonics in real time for the standard CPU. As a VSTi. I'm going to go try it.

Uh-oh. Think it will be raining Fredholm kernels from Banach space, any minute now. Best bring in the herd.

That's hilarious!!

Best watch out for successes where you least expect them.