it's better power wise as well; it's a bit complex to explain, but
it's better to execute the code you need to execute at full speed, and
then really quickly go idle, than it is to execute at a much lower speed.

Maybe a simple example (I plucked these numbers out of the air, they
don't represent any real cpu that exists) will help:
Say you have a cpu that consumes 40 Watts at full speed, and 30 Watts
at half speed, and 4 Watts when idle.
You have something to do, lets say mp3 decoding of 1 second of audio,
and that takes a full second at half speed, and one second at full speed.

At full speed decode + idle, that is half a second at 40 watts (20
Joules) and half a second at 4 Watts (2 Joules); total is 22 Joules.
At half speed decode, that is a full second at 30 Watts, so 30 Joules.

So, what ondemand does would cost 22 Joules, while a "hit the exact
frequency" governer would cost you 30 Joules.....