One of the most über-cool things I've seen in a while: a Playstation 2-based computing cluster at the NCSA. They've installed Linux on the thing, and have harnessed the vector units, intended for graphics processing, to crunch through scientific computations (which are also quite vector and matrix-heavy) quickly.
You have to admit, the cost/compute power ratio of game consoles these days is quite something. And it's not just on the console world, but the graphics cards on PCs as well.
Since 3-D graphics essentially is comprised manupulating vectors very quickly, the Graphics processors found not only in the PS2, but the latest PC graphics cards are now essentially very fast stream-based vector processors, and can be readily harnessed for general-purpose scientific computation other than graphics: particle, cloth, fluid simulations. The GPU replaces the CPU for computation, and texture or other video memory, with its much higher bandwidth and lower latency than system ram, is used as a backing store for data.
A lot of the GDC and SIGGRAPH 2003 papers focus not on graphics directly, but on scientific computations using the CPU. It's very cool, and if nVidia and ATI the like ever want to expand into a new market, they should build cards with multiple GPUs each, and sell them to the scientific community, or to non-realtime CG places like Pixar to accelerate their offline rendering.
This page has a good summary of research in this area: http://wwwx.cs.unc.edu/~harrism/gpgpu/index.shtml
Maybe I should go back to grad school after all...