These toys can be programmed in C language (well, very similar to C actually) for tasks not only related to video applications.
As example the latest NVIDIA GPU, the GTX 280 sports 240 processors!!
Each processor runs at 1296 MHz and memory bandwidth is an incredible 141.7 GB/sec.
What is most important each processor is programmable in C with so called CUDA environment (see http://www.nvidia.com/object/cuda_learn.html)
This thread title suggests Monte Carlo analysis mostly because it is a technique well suited for parallel processing being each "little game" independent from each other.
So the advantage of a chess engine implemented on a GPU are
- Massively multi parallel: row power order of magnitudes higher then any octal or even bigger traditional box
- Programmable in C
- Low cost, easily affordable
- Evolution of GPU has been much faster then CPU in the past and the trend is still that: GPU market (video games driven) is far more important then quad or octal big boxes one (mainly enthusiast driven)
Finally the question

As anyone ever tried the GPU way to develop a chess engine ?
Thanks
Marco