Easy-peasy! Raytracing's a pretty simple procedure requiring only a little code, which'll fit neatly into LS's. The scenery data was generated on the fly from 2 bitmaps, which is a regular data structure that can easily be double/triple buffered with DMA's to ensure the SPE's always have code to work on. This means all the code and data is directly available on fast LS without the latency incurred by needing to work along with L2 cache. Even though the FP of the Altivec may well be comparable with the FP unit of the SPE clock for clock, it's keeping the units fed with data that SPE's are good at when the data accesses are pretty straightforward.
An interesting comparison to make would be the PPC970 versus a SPE on a typical raytracing scene, of lots of objects and more erratic data fetching. This should see SPE's advantages reduced and the attained performance be a lot less than theoretical max.
As XBD's already suggested, take a peeps at this independant companies results using Cell
http://www.mc.com/cell/demo.cfm. In a task very well suited to Cell's architecture Mercury Computer Systems are getting 100x the performance that they were getting from a 'high end server core'. In contrast, Alias' cloth simulation was only getting a few times the performance of a P4 from Cell, so not every application is going to be much, much faster on cell, some'll be worse, but, strangely, those tasks that Cell was designed to do well at, it does well at!
