I don't deny that something may be happening at the system level, my comments are regarding the CPU only. I have been working at a CPU designer for Intel for over 10 years starting with the Pentium and more recently on the Itanium family. I have spent most of my days in the last year determining the critical paths on a 90nm CPU ("critical paths" are the circuits that limit the frequency of a design) and finding fixes for them. I spend a good chunk of my day running shmoos (a 2D plot of voltage versus frequency used to visually help with determining circuit marginalities) and trying to isolate issues.
I have never seen a mechanism that causes a given CPU operating at a given V/T point to speed up occur on a statistically large sample of parts. Integrated circuits are fundamentally crystals that are as perfect as the manufacturer can make them when they come out of the fab. They are - to the limits of manufacturability - as perfect crystals as we can make with as few dislocations, voids, grain boundaries, vacancies and impurities as possible. Using these integrated circuits involves subjecting the crystals to sharp thermal changes as well as voltage changes that result in the crystal gradually degrading over time. Normal operation changes the crystalline structure over time for the worse. Carriers get stuck where you don't want them. Charge traps are formed at the boundary. High current wires that are thinned and under thermal stress thin even further and grain boundaries become more pronounced. Under normal operation, eventually the integrated circuit will stop working. IC's are not like cars or other things with mechanical parts. There is nothing to "wear in". They only wear out.
I am not saying that people who claim to see some speed-up from some form of burn-in type operation are lying or are wrong. Just that - based on my experience - what they are seeing is not coming from the CPU and that I'm not sure what they are seeing.
I had a lot of reservations about the overvolting advice that was going around the net several years back. The idea was to run the CPU at an elevated voltage for a few days, and then afterwards it would run faster - and this idea is just fundamentally wrong. There may be some truth to the idea of setting thermal compound - or maybe it sets the grease in the fan ball-bearings or something - but the idea that the CPU is running faster was just plain wrong. And the idea that you could bias a PMOS gate such that it is repelling electrons and then somehow have a bunch get trapped in the gate creating trapped negative charge carrier is so wrong that I could show anyone in this thread why it's crazy with a few diagrams.
I don't have any such reservations with the idea of undervolting. If people think that it helps - it might in some way that I'm not familiar with - then it's certainly not doing any harm. I don't pretend to know all of the answers - I can only relate my experience. I have never said that "burn-in" doesn't do anything - only that whatever is going on, it's external to the CPU.
It's also worth mentioning that it's easy to be fooled by 'noise'. A very small improvement, may not actually be an improvement at all. There's a great (great!) book called "Voodoo Science: the road from foolishness to fraud" by Robert L. Park which talks about people throughout history who have been fooled by the "placebo effect", and by people who have been fooled by "noise" into thinking that something exists where it isn't (cold fusion, for example). 25MHz, or even 50MHz is a very small amount of 2000MHz or 3000MHz. I have a hard time reliably and repeatably isolating my failures down at that resolution using testers, thermal systems and power supplies that cost in the high six figure range. Just as I try to keep an open-mind about improvements, it's worth people keeping an open mind regarding seeing possible low-value illusionary "signals" amongst the noise.
(
link to "Voodoo Science" at Amazon.Com Check out the reviews at the bottom... I'm not the only person who really liked it. If you like reading well-written and entertaining books on science, this one is a winner)
Facebook
Twitter