Ok, here's the deal with "MHz myth". Read carefully, as I won't repeat twice.
The CPU clock in MHz is an indication of how fast the CPU can move, so to speak. MHz is usually a very good indicator of speed _only_for_the_same_CPU_family_. Again, only for the same CPU family. G3/G4/G5 are in the same family. CPU clock among those processors are a good indicator of relative speed. Same goes for Intel P2/P3/P4. I said "usually", because even in the same family, the performance can be improved drastically by the use of more pipelining, branch prediction, superscalar, etc, which is not the scope of this article.
The CPU clock is a VERY poor indicator of performance across different processor family. 1GHz G5 and 2GHz P4 can't be compared with just clock rate. The reason for that is the "work" that each processor can do per clock cycle is different. I said earlier that the clock rate is how fast the processor can move. Guess what, each processor have different "efficiency", so to speak. For example, to perform a simple arithmetic multiply may take one processor 1 clock cycle, while another processor 10 clock cycles. Yet the third processor may perform addition and muptiply in one clock cycle! That's why you can't just compare MHz vs MHz, because the work that the processors can perform in each MHz are different.
32bit vs 64bit processor, IMHO, is a big marketing hype. 64bit processors are not twice faster than 32bit processors as they might (wrongly) imply. The advantage of 64bit processor is what Avalon has explained. In order to add two 64bit numbers, it takes one clock cycle for 64bit processor, while it takes several (not just two) clock cycles for 32bit processors. But then, we don't generally deal with 64bit numbers. The maximum addressable number in 32bit space is 2 (or 4) billions. That's certainly big enough for me, although certain applications may need larger address space.
Having said all these, I'm not surprised to find 16% increase or whatever with iMac G5. Look at their benchmark. One is completely disk IO bound, which has very little to do with processing power, and the other two are completely CPU bound. They are not "good" benchmarks because they are biased towards certain subsystem, rather than measuring overall system performance.
