Re: Overclocking stability with gaussian98

Thu, 15 Jul 1999 16:15:21 -0500

From: Christoph Wasshuber <>

Let me put my two cents in this overclocking discussion. Since I am working in silicon technology R&D I can give you some insights from the fab. TI is not manufacturing standard CPUs but we do quite good with fast DSPs, so the experience should apply in general.

Overclocking will reduce the lifetime of your processor. But since the lifetime of CMOS is very hard to control precisely, that might be or might not be a problem. It is very difficult because it changes from lot to lot, meaning if you buy two processors at the same day from the same distributor or vendor you might get two totally differently behaving chips, with respect to lifetime. The minimum lifetime of CMOS is somewhere around 10 years. At least that is the target of manufacturing. But again that is a statistical value, so most of the shipped chips do much better. That means that even heavy overclocking should not realy matter over the usual lifetime of a PC.

Going to finer and finer structures 0.18um, 0.13um means that the production facilities are stressed to their limits. Which means that variations become larger because smaller transistors mean also more sensitivity to minor process changes. Chips will have larger variations than in the past. So you can end up with a much better performing CPU than the specs tell you.

Stability is another issue. The best would be to write a program which tests all possible signal combinations in the critical parts of a processor. Running an application, no matter how difficult it appears, is not a reliable general test. Of course, if all you do is run one application it is enough to see if this application runs stable. However, it is enough for one line of assembly code to change, to let the CPU produce errors. Todays CPUs are pretty sofisticated. So a simple black box test is not reliable as a general test.

Nevertheless, I think overclocking is a smart way to get more for less.