If you're talking about pure raw power, the Cell wins. To back up this fact, this page shows that the Cell is capable, at least under certain circumstances (in this case, an average double precision Linpack 4096x4096 matrix), of attaining performance equivalent to 100 GFLOPS in a real-life situation. In contrast, the Xenon's theoretical peak performance is only 115.2 GFLOPS.
However, if you're talking about how much can be realistically utilized using today's dev tools and programming languages, well, I'm not really sure. I think the Cell is more powerful by a small amount by itself, but I don't know enough about programming for multiple processors to fashion an informed answer. I do know that in a video game console it can give programmers an advantage by using the SPEs to take some of the "grunt work" (such as physics calculations) off the GPU.
Not trying to be a fanboy. Of course, it's hard when you own the best console eve... dang it
