let me explain how processors are manufactured.
IBM makes a disk of cpu's and predicts the quality of the processors, they used to hand speed bin them but it is very uneconomic to do it that way so they have to hand speed bin a few and then use the data from the hand binning to predict the quality of the processors, the cpu's on the outer edge of the disk are the worst quality and the cpu's in the middle are the best due to the silicone crystalline structure, their are two limits to the clock speed of these prossessors, the first is the physical length of the longest consecutive string of transistors, and that is usually the number of transistors in the execution unit divided by the pipeline length, though that data is only known to IBM (unless i did some tedious maths and studies die diagrams which tbh i cant be arsed to so right now) you then work out the time it takes the a signal to cross that distance (say 0.1m divided by 3x10^8m/s) divide one by that number and you have the maximum attainable clock speed in Hz, for the 90nm G5 i suspect it's about 3.2GHz.
the other limit is the voltage, signals do not look like the digital signals you imagine they are more like sin waves, the constant is the voltage required to register logic level 1 or 0 and if that is the same as the maximum supply voltage then the signal will only register at the top of the wave and there for will be at logic level 1 for a very short time, when the supply voltage is allot higher than the voltage required the transistor will be triggered much faster as it at the steep part of the curve and the signal will be allot clearer, and can be clocked allot higher without becoming unstable, so IBM needs to set the highest voltage the the processor can take while remaining stable, as the cpu's around the edge are more susceptible to electron tunneling and electromigration and they shorten the life of the processor as they eat away at the cpu and make let signals jump track, and high voltages make it worse, so IBM rates the cpu lifespan then applies the highest voltage that will comply with that then rates the clock speed according to the voltage, this is all predicted every time their is a respin (change in design/technique/tollerences)
IBM cannot overclock by the definition of the word, they can be too lax in their clock speed ratings, but we dont see widespread reports of G5 cpu failure do we?