The only safe limit is the one set by the manufacturer. Manufacturers set it to the highest value that they can while staying within the limits of their reliability goals.
It all depends on cooling since for most purposes, it's the heat from the voltage that endangers the CPU before the voltage itself does.
This is not true. It's the other way around. There's a linear dependence on mean-time to fail (MTTF) on a CMOS part to temperature, there's a square dependence on voltage to MTTF.
In non-statistics speak, increasing the temperature a bit will make your chip a little more likely to die. Increasing the voltage a little bit will have a much bigger effect in killing your chip.
In the "old days" - prior to 0.18um process technology - the dominant reiability failure mechanism was electromigration. In electromigration, higher temperatures and higher voltages reduce the average time to failure. Because electromigration is dependent on current density, increasing voltage is worse.
Once the industry switched to dual-damascene copper technology, electromigration failures were relegated back to design-related mistakes. Instead the dominant failure mechanisms became time-depenedent dielectric breakdown (TDDB), PMOS BTI (although this is pretty much accounted for in manufacturer burn-in), and NMOS hot-electron gate-impact ionization (NMOS hot-e). In all three of these small increases in voltage can result in large reductions in operational lifetime. In fact, in the case of hot-e, it gets worse with lowered temperature.
For more details, you can read through this (rather condensed) notes page
http://www.eie.polyu.edu.hk/~ensurya/lect_notes/Reli_Fail/Reli_Fail_notes.htm. Although the whole page is an interesting reference, the part relevant to this discussion starts with "Properties of Metal-Oxide Silicon (MOS) System".
Trust me, you will not kill your Cpu with 1.550v...
Clearly increasing the voltage reduces the MTTF from the manufacturer's ratings, so I don't understand how you can state this. How large is the sample size that you are basing this statement on? What data are you using?
Patrick Mahoney
Senior Microprocessor Design Engineer
Intel Corp.