> So basically it's "a dnetc client thing" after all?

I'd say dnetc and probably many more that's tuned (per compiler switch or assembly optimization) to any of 60x/7xx/74xx microarchitectures, which are closely related (603e -> 740/750 -> 7400), whereas the 970 microarchitecture is a different beast (604e -> POWER3 -> POWER4 -> 970).
We have this kind of issue on MorphOS also with E-UAE for instance, where a special G5 build has been created because the generic PPC build runs too slow on G5. I guess there's a huge share of MorphOS software affected by this general issue, but it's mostly only noticeable with performance-critical software like benchmark tests or emulators.

> As for the AltiVec computations, here are lame_vmx results: [...]

Per-clock performance normalized to "1.00 = slowest" is this:

1.48 : PMG5 PCIe
1.44 : PMG5 AGP
1.34 : iMac G5

1.00 : PowerBook

1.07 : iBook

Interesting. That's not what I would have expected.