I'm also very intrigued by this topic so I thought I'd do some tests. I used Speedometer 4.02 which is a Mac benchmark program. I tested ShapeShifter 3.11 with different graphics drivers(P96 internal, external and CardTrickEVD) and Fusion 3.2 in two resolutions(640x480 and 800x600) and two colour depths(8 and 15 bit). The test is called Color Quickdraw.
Speedometer 4.02 compares the results to a Mac Quadra 605; all results for this machine would be 1.0.
The results are a long list of boring numbers so I'll just give the conclusion. (Also I'm too lazy to make a table for it;-))
P96 internal is by far the fastest; about 1.8 times compared to Quadra 605 in all tests. P96 external is about 0.5; so is CardTrick. Fusion is about 1.5...
I also did some CPU/FPU tests; ShapeShifter and Fusion are about equal with about 3 compared to Quadra 605.
These results/numbers are for my system config as it stands in my signature.