I'm curious about this too- One way I thought of doing this would be to estimate it via experimentation. Find some calculation task or program (like multiply two 10000x10000 matrices) that is basically totally CPU dependent, and measure the time it takes to do it on highest, auto, and low. the times should roughly be proportional to the CPU frequency, and since you know that highest is YYY GHz (whichever model you have) you can scale it from there. I know of no such program so far though- although it would be easy to write, for example in Matlab.
Ruahrc