There is no easy answer. My approach would be something like this: * Take 100 identical machines and run them with a specific setting for a year. Take the numb

There is no easy answer. My approach would be something like this: * Take 100 identical machines and run them with a specific setting for a year. Take the number of failures and write it down. * Take another 100 machines and run them with a different settings for a year. Write down the failures. * Do this for like 10 different speed settings. (You need 1000 machines and have to wait one year of time.) * Make a graph of the data. Log of failures over linear speed axis. * Try to see a curve in the graphed data. * Determine the failure rate that you want to achieve. * Look up the speed setting in the curve. * Now you have your speed. I guess only people like the processor manufacturers actually do it this way. Most other folks just do guesswork. If you have advanced knowledge about the failure curve you might come up with a law like this: * Turn up the speed until you hit 1 failure per 24 hours under heavy load. * Turn down the speed by 30% and sell.