baka on 23 February 2007
The Cell is a nice architecture, but the SPUs are basically just asynchronous co-processors with their own RAM and slow transfer times to main memory. It's an excellent architecture for offloading small tasks from the main PowerPC core, but you need tasks which are highly parallelizeable, not dependent on more than single precision floating point unless you want a performance hit, can be accomplished in 256k, and predictable in runtime in order to take proper advantage of them. Threading in the classical sense is not terribly practical since it depends on using memory in-place, so you really need separate tasks that provide results VERY quickly to make it work in a video game.
