Just a really quick follow-up. Using pkgsrc/benchmarks/flops (which 
specifically says that it's small enough for any cache), I did a few 
comparisons. For simplicity, I am using units of mflops per 10 MHz since 
the clock speeds are all over the place. Also, flops includes four 
results; I'm using results 1 and 4, since 1 includes fdiv and 4 does not.

G5:
3.3, 3.42  (OS X)

G4 7457:
1.95, 3.39  (NetBSD 2.0)
1.96, 3.53  (OS X)

G3 750FX:
2.27, 4.94  (OS X)

G3 750:
2.11, 4.87  (NetBSD 1.6)
2.11, 4.73  (NetBSD 2.0)

604ev:
2.4, 6.05   (NetBSD 1.6)

603e:
1.94, 4.2   (NetBSD 1.6)
2.03, 3.91  (NetBSD 2.0)

m68060:
1.3, 2.6    (NetBSD 1.6)

m68040:
.88, 1.16   (NetBSD 1.6)

So yes, the 604 beats out everything except the G5, and even beats out the 
G5 in floating point code without fdiv.

John