Thanks to a pointer from SlashDot, I found this article by NASA Langley Research Center researcher Craig Hunter.

The study compares the G5 against a P4 (and a pair of G4s) in a single processor configuration (done by turning off the second CPU using Apple's CHUD tools) on scalar operations using the Jet3D application (which, it appears, is written in Fortran).

The conclusion is that scalar fixed-point in the PowerPC is finally back above the per-megahertz ratings from Intel. The author notes that the P4's are (of course) available in higher Megahertz ratings and therefore can still outperform the G5 when the highest speed CPUs are compared. However, the G5 at 2GHz appears to outperform the P4 at 2.66Ghz, even without optimizing compilers for the G5.