It's a good question, and one that comes up here a lot and that there's not generally a good consensus answer.
There are plenty of low-level reasons for why a chip will stop working, and there are equations which cover the statistical probability of failure will tell you what the key variables are that govern that failure. In some failure mechanisms, temperature and voltage play an equal role... in others the voltage dominates. In one failure mechanism, the cooler the CPU the worse it is. But it's hard to know what the specific mechanism that will govern the reliability of a microprocessor on a given process technology unless you happen to be involved in reliability studies for that particular product. And even if they did know, it's still a statistical thing.
For example, for a long time CPU wiring inside the CPU was made of aluminum with some copper and silicon atoms thrown in. Aluminum atoms are fairly lightweight and so when a huge number of electrons are flowing down a wire, they can actually push the aluminum atoms around. It's kind of like trying to move a car by firing ping-pong balls at it. It's a crazy idea unless you have a ridiculous amount of them and they are moving very fast. If you move enough metal atoms down the wire, the wire will start to have holes in it, and this will increase the temperature of the wire, which makes things even worse, and eventually the wire will fail. This is called electromigration and it was the dominant failure mechanism on CPU's up until about the 180nm process technology when copper metal started to get used to make the wires. Copper atoms are much larger and they pretty much stopped electromigration from being a problem for a while... although it's starting to become an issue again.
Electromigration is described by "Black's Equation" which is: MTTF = A*/(J^n)*e^(Ea/kT). MTTF is "mean time to failure" and is the statical time it takes before the wires fails, A is the cross-sectional area of the wire, J is current, n is a scaling value, Ea is the activation energy which is dependent on the metal, T is the temperature and k is the Boltzman constant. So for a given wire, the area, the constants, and the activation energy are all fixed. So the two that you can mess with are the current density and the temperature. So MTTF is directly proportional to the square of 1/J, the current density, while MTTF is directly proportional to e^(1/T) - in this case voltage is much worse than temperature. See also:
http://en.wikipedia.org/wiki/Electromigration .
But electromigration ceased to be much of an issue in pretty much the 7-8 years in the industry and now there are plenty of new ways for CPU's to die. Other things that will kill a CPU nowadays are: gate-oxide wear, hot-e gate ionization, oxide defect generation, PMOS negative bias temperature instability, and several others. In the case of gate-oxide wear, there's a temperature acceleration factor that gives a higher dependence on elevated temperatures than on increased voltage which has a linear dependence. While Googling for this post, I came across this paper (
http://www.ece.rutgers.edu/~kp...apers/NanoSymp2001.doc ) which discussed a variety of oxide failure models and has the conclusion: "The understanding of thin gate-oxide wear out and breakdown is far from complete. None of the model in the literature is without shortcomings."
So in a nutshell, even the experts in the industry don't know. If you want my advice, though, be very careful of the voltage, don't press your luck too far and I wouldn't recommend listening to anyone who says "voltage doesn't really matter - it's the temperature that will kill your CPU" because the academic literature doesn't support that at all.
Facebook
Twitter