Effect of high Vcore and electromigration on expected failure time for Tbred B/Barton
In the last post, I estimated and related the CPU over voltage with expected 50% sample failure time (life expectancy) based on electromigration emprical results (Black's equation). Here put them into numbers for CPU voltages.
Statistically, for the same level of temperature specification,
A 10% increase in Vcore, would shorten the failure time to 83% of nominal failure time.
A 20% increase in Vcore, would shorten the failure time to 69% of nominal failure time.
A 30% increase in Vcore, would shorten the failure time to 59% of nominal failure time.
A 50% increase in Vcore, would shorten the failure time to 44% of nominal failure time.
So a 30% increase of Vcore reduces the 50% sample failure time to 59%. 30% over stock voltage for Tbred B/Barton are
- 1.95 V for DLT3C, such as the famous Tbred B 1700+/1800+
- 2.08 V for DUT3C, such as the popular 2100+
- 2.15 V for DKT3C, such as the Barton 2500+ or higher.
E.g. If the nominal CPU life expectancy is 10 years, for Tbred B DLT3C
- 30% overvolt to 1.95 V, the number would be down to about 6 years (59%).
- 20% overvolt to 1.80 V, the number would be down to about 7 years (69%).
- 10% overvolt to 1.65 V, the number would be down to about 8.3 years (83%).
They seem to fit nicely w/ the AMD absolute rating 2.05, 2.15 and 2.20 V on Vcore for the DLT3C, DUT3C and DKT3C respectively.
Max Vcore for Tbred B and Barton (page 5)
Based on the analysis, we can rule out the guessing numbers of 1.8 - 2.0 - 2.2 V for max Vcore flowing around and also the concern of failure within weeks or months.
As far as
temperature to not having additional adverse effect on chip behavior from electromigration on top of voltage, it should be below the max temperature rating of 85/90 C (for TBred B/Barton). So using a temperature cap of 65-70 C is reasonable, since above which most CPU would be overclocked above the break-even point of 10 MHz/C for Tbred B and Barton. Further increase in voltage and temperature beyond 30% and 65 C, even if it is stable, one would get very little return in MHz, but greatly shortening the expected failure time. (Besides temperature is kept under 65-70 C, HSF, motherboard FSB, memory, PSU, ... are assumed not to be limiting the stablity of the system.)
In conjuction with the MHz gain from 10%, 20%, 30% over voltage, one can pick and chose the tradeoff between MHz gain and the reduction of statistical expectancy of CPU failure time.
If one plans to use the CPU for 20 years, or if one is not comfortable of using a 30% higher voltage at which CPU is working above the break-even point (10MHz/C) of frequency and temperature (on air), one would not lose too much MHz even the Vcore is lowered by 10% (~150 mV) at that level, estimated by about 100 MHz. For practical reason, apart from short term benchmarking and fun, trading 100 MHz for 150 mV lower in Vcore is justified.
There comes a point when they refine the process so well that they really can't make chips slow enough. It sounds odd but it really is true. A great number of things in the world of electronics are made like this. Oh this widget doesn't cut the mustard to be the real expensive model of widget so we will badge it as a cheapo model. What happens when you get so good at making widgets that everything starts passing the test? 
