My findings were that I hate bus errors. :) Actually, as you may know, there was a shooting at VT in April, which cut the semester short. The Cell project was therefore optional. Myself and one other student were the only ones who asked for access to the cluster, and I don't think he ever coded anything. But because I only worked on it for about a week, I didn't finish it. I did get software pipelining and partial double buffering going, but I needed to reorganize the shared data access in order to get all 6 SPEs able to do double buffering while sharing the matrix data. I expect I would've ended up in the 50-75 GFLOPS range when all was said and done.
A raytracer is one of the things I'd like to write if I get my hands on the Cell again. It's never going to be usable in a game, unfortunately. Complex scene render times are measured in minutes on fast CPUs, and an order of magnitude faster than that is still more like 30 seconds per frame than 30 frames per second. But I would say that if Pixar doesn't have a rendering farm made of a few dozen PS3s, they're probably doing something wrong.







