GPU for computing

[pegasis]

How does one add GPU (multi core) for super computing on an existing PC and tap into the computing power on the GPU?

is this done with some special software or some special hardware to enable the GPU to be used for computing instead on the Chips on the MB??

 

[BrightCandle]

Special software. There are three major APIs programmers can use to access the compute capabilities of the GPU, openCL, CUDA and DirectCompute. The software needs to be completely rewritten to work on the GPU, its a very different approach to concurrency than most approaches done on CPUs (which are usually thread based on SIMD based).

If software doesn't support it there is no way to make it work, you need software specially written to make use of the GPU, and much of the software today does so very poorly because they aren't great fits for that type of parallel machine.

 

[easp]

Adding to what BrightCandle said: some apps are already coded to take advantage of GPU when available. For example, I know some Adobe apps and other media apps, do, as do some engineering apps.

OSX provides some frameworks/libraries that make use of the GPU, and even provide a way to have some code run on the main CPU AND the GPU. Even so, there is a lot that doesn't yield easily to being run on the GPU. Apple 's latest MacPro is a GPU monster, with one GPU pretty much dedicated to computational tasks. To show it off, they released a new version of Final Cut Pro, their video editing app. The GPU provides a huge speedup for some operation, very little for others.

I'm less familiar with the windows side, but I expect similar issues exist.

Some code is easy/worthwhile to run in parallel, but a lot isn't. Of the code that yields to parallelization, some of it yields to GPUs, but pretty much all of it is going to be easier to do on a multicore CPU.