Cuda is a from of parallel computing from NVIdia which offers processors and threads of C computer power on graphics cards which sends quite a few CPU's and even mild FPGA processing solutions potentially crying for their mothers (or so) when it concerns possible MIPS/FLOPS power.

As it goes with parallel computers, including most moderns supercomputer farms, the actually achieved effectiveness of the programs running on them doing some usefull computation work depends on how well the problems lends itself to parallelisation. And how well the latency and bandwidth of the parallel grid behave in practice.

I've used Cuda a bit lately, not on a desirable I7/GTX260 but on a 3GHz Pentium D with a 512MB GF9500GT Nvidia graphics card (16x PCIexpress), and since all the examples compiled great on the Fedora 10/64bit OS I used for it (gcc 4.3 I believe) with Cuda 2.1, I thought I'd try to use the new toy a bit and see what I can do with it, so I recently made a Cinepaint plugin with it and have sucessfully experimented with a cuda extended tcl/tk interpreter.

I wouldn't say a beauty price would be in place, but  things can work, and the specs of even a mild cuda card (compared with a whopping 4 Tera ops/sec and double precision floats from a Tesla rack) are such that apart from the advantage of a processor seperate from the CPU running intensive computations and the fun of using powerfull new technology with other limiting factors than standard CPUS, the actual raw computing power and not to forget practical memory bandwidth are quite attractive compared to even really fast PCs, depending on the type of algorithm being used. And the cards certainly aren't expensive.

  A Test Setup

   I want to try out a standard setup of an audio and midi input and audio output to a Linux program (on Fedora) with Jack and Alsa, and see if with bearable latency the audio data can be transfered from and to the Cuda processors on the graphics card.

I made a test program with a cuda routine called together with the audio frame processing, which is an example cuda routine kernel + IO of arrays, and it takes about 1.3mS to communicate and execute, so should run well in conjunction with a 256 sample buffer at 44.1kHz.

In practice the program runs, and I'm working on linking the cuda part in the audio loop. First I want it to not skip frames more than very few times only, by making the CPU audio loop empty and working on the way the various procedures are called (maybe a jack library compiled by nvcc) . My currentl method can work reasonable with a select() call, I think, but even then it probably isn't really up to much multitasking: it skips frames every now and then, though maybe I can get it to behave, I spent one night only, and didn't do much special (yet).

I've prepared including the whole bwise blocks with maxima formulas to compiled program to be used also with cuda. A working fortran to C converter which I can run can probably work with Cuda, and that seems liek a nice idea. In principle it could even run on a webserver! For instance to generate audio waves from formulas or block connections in a short time, and make good graphics from them, by parallel processing.

   Purpose

    Probably I'll want to make my formula rendering software work on cuda or anything that would seem reasonble to take advantage of the not soo much chache troubled and quite powerfull parallel processor. Including the (closed source) surround reverb I made already some time ago.

Clearly my string simulator lives almost in a class of it's own computation intensity wise, and it could even seem the architecture is a but made for it's type of computations...

Also, power audio graphics can be made, cuda can directly talk to openGL for high bandwidth and high resultion graphs and interfaces.

  Interested People

   Probably this won't be an Free and Open Source project because I don't believe that's reasonable anymore, since no work, money, or kudaas have been sent to me, and I'm not working on the subject of making myself only free and open sourced! I still do believe in the thought, and so of course free software and open ideas will remain and some new will be made, but this subject is not.

   Scientific interest people may have in it, so I'll describe in limited but usable detail some ideas I use. Probably people will not mention me as motivation or effectively as the engineer who gave them those, and take them without acknowledging me, but they'll find the whole setup cannot just be coped, it's to complicated for that and God and homeland, and many free and self-working people and engineers will catch on and hopefully rock the whole of the patent suckers anyhow. At least I can look in the mirror every morning and not dispise myself for being a dreadfull and pityfull human being, and I continue to believe and support the work and ideas of the FSF and others (Like Fedora and the people who made the Cuda environment work for free).