Modern GPUs are increasing in programmability and these chips can do more than just graphical computations. They can now be used as a coprocessor, and they can be integrated for a set of tasks. GPGPU (General-Purpose compuation on GPUs) is such an initiative that contains a catalog where the GPU can be used for general-purpose computation.

The big challenge, is to translate the everyday applications to two-dimensional graphic functions, like texture mapping. In other words: Pretend that everything is a game (source). 

As an example in this article and results, a quicksort algorithm of 18 million records in Visual C++ took 21 seconds, while the GPU took 2 seconds! What are the results for a Quad SLI setup?

Microsoft research is apparently working on a system that simplifies the programming of GPU to general-purpose tasks, it's called Accelerator (simplified programming of graphics processing units for general-purpose uses via data-parallelism).