I'd like to first note that I prefer gaming on the PC. But...
The fact that performance of your average PC game relies so heavily on your GPU is a result of the algorithm choices those game designers made much more so than many think. They make these choices because no assumptions can be made about relative strengths of CPU and GPU, bus speed, etc. This is the way things have been for a little while now in the X86 commodity hardware world and I get the impression that this is leading people to think that this is some iron law of the universe.
Consoles have a significant advantage simply because their hardware is invariant. Invariants make problems simpler. This is something that transcends computers, any scientist will look for invariants -- or artificially add them -- in a system to simplify his model of that system.
At the very least, the first thing that one usually tries to impress upon students in any undergraduate computer architecture class is that nearly any of these "more powerful than thou" statements put out by companies are full of lies of omission.
While I am starting to change my opinion of GPGPUs strength relative to FPGAs, I have my reservations about them relative to traditional CPUs (cell and it's SPEs still fall into this category). Stream processors in general are only good for non-branchy, trivially parallelizable tasks, with very small memory windows. These things are why traditionally most of the compute time for volumetric particle rendering was in the cpu. There are now algorithms for the new cards, but they make all kinds of assumptions due to the limits of the architecture. The last part is extremely important in something like a sparse matrix algorithms (which are normally faster if you know your matrices are sparse) or even decent, quick AA.
There is alot more, but hopefully my point that things are not as simple as they seem gets across.
