In a comment to one of my previous posts, "TimC" asks if the performance of my simple raytracer is "anywhere near realtime".
At the moment it's certainly not, although there is definitely the potential for it.
I haven't spent much effort on optimising the ray tracing algorithm, because I'm more interested in optimising what I've got, and the issues I come across that are interesting and different because I'm using the Cell. So at the moment I have no spatial partitioning at all, I just do a brute force raytrace per pixel. Even a simple bounding box would significantly speed up the current implementation, especially for the test scenes I'm using where there's a lot of empty space.
Some IBM guys in the US have been working on a serious raytracer, they call the "Interactive Ray Tracer", it is seriously impressive. They have a cluster version running on three PS3s. From memory their test scene has something like 300,000 triangles, compared to mine which is about 30 or less.