Errr. If the Cell is putting stuff in the RSX memory for the RSX to work with, I'm not sure why the RSX has to be too "tapped out." If the Cell is computing all of the geometry, it can write that information right into the RSX (and do it fast), and have it waiting there for the RSX to start doing all of its shader routines/rendering/whatever.
This is what I'm curious about with respect to using the Floating Point power of the Cell to calculate Random Variable distributions. Maybe we no longer need to abstract the decision making in order to reduce the floating point instructions, since Cell has a lot of floating point power.
Branching is important. But even then, determining which branch to take requires some sort of metric by which they make their decision (where I'm curious if the floating point power could be an advantage, as stated above). I know there's a lot of hoopla about the poor branch prediction of the Cell, though I'm not sure how (this is starting to go beyond the scope of my studies). According to IBM:
Also, I have heard concerns that the Cell, while powerful, is inappropriate as a gaming processor. I find this a little unusual, as much of IBM's articles seem to indicate that it is probably best suited for a gaming processor.
Meh.