General-Purpose graphics processor

The general idea

Graphics processing units (GPUs) have very fast memory interfaces and, in higher-end models, a very large amount of parallelism. This makes them tempting targets for implementing compute-intensive tasks on the cheap.

However, they are not designed to be programmed with a general-purpose workload. Also, rendering 3D graphics is not accuracy-critical (the odd incorrect pixel is not noticeable within limits) so the floating-point implementation may lack accuracy compared with a conventional processor.

The challenge

Understanding the programming model as well as limitations of the instruction set will be interesting. Issues you will need to consider include

focussing on one specific range of GPUs where documentation is available
understanding the programming model and its limitations
analysing the potential for performance gains taking into account extra overheads like communication with the GPU
working through an example of significant size to clarify the issues

The programming tools in this area are a work in progress so a complete working implementation while desirable is not essential, so be prepared to do an analytical evaluation.