Since folks are having some stability issues, I'll just offer my 2¢ from my stability work last night. The system was basically stable, but I noticed a couple of app crashes which made me want to really nail it.
First, I had a problem where Prime95 wouldn't say I was stable no matter what. I was running custom 20K FFTs, default memory, and it would have rounding errors right away. Keeping Vcore high I dropped 100MHz, no improvement. Dropped another 100, same problem. Switch to "small FFT" or "blend" and it works fine for 45min+. So I suspect this is a bug with Prime95. I'm not sure if that's the case, but I was definitely barking up the wrong tree.
Next I used Aida64. Since I was convinced my FPU was OK from Prime95, I focused on the other tests it provides, "CPU," "Cache," and "RAM." Running any of them alone didn't give any problems after 10-15min, but running "cache" and "RAM" together would cause a hardware failure to be detected.
After trying a bunch of different voltages, I found some setting that work with my system, and my be illustrative. First, I manually set the system and memory controller voltages to their default values. Having them anything else made my system less stable. The big discovery was to set Vpll as high as reasonable
, which for me was 1.875. (Default is 1.8, and Intel recommended max is 1.89V).
This is in some contradiction with many guides, which suggest lowering PLL voltage to reduce heat. The higher PLL voltage may increase heat, but it crushed my memory/cache bug.
On a philosophical note, I'm not big on stability, especially as this site conceives it. My belief is that there are enough bugs in the software that worrying too much about processor stability is misguided. The exception is for very intese technical computing, like folding. Then being Prime Stable for 12+ hours matters. But all my technical computing runs quickly enough that I can do it twice and make sure the answers are the same
The other (and bigger) problem with the "Prime Stability" idea is that, as my experience shows, it's not a complete test of the system. It heats the processor, which is important, and tests the FPU. But all the integer math, the crypto, random number generator, cache, graphics, memory and system controllers, none of these are explicitly tested by Prime95, Linpack, or other FPU tests. If I hadn't had Aida there to check other parts of the system, it would have never occurred to me that the PLL was the issue.
Aside from Aida, my favorite ways of stability testing are playing games and opening a bunch of tabs in Chrome. Of course, Prime is a wonderful tool. But the standard way of thinking, that "errors in Prime mean you need to boost Vcore and/or reduce clocks" was wrong.
And like I said, my main criterion for stability is... whether or not my system throws errors.
But it's a personal thing, and I just want to encourage folks to think flexibly about it.
EDIT: Of course, after posting this diatribe I have a game crash on me. Went back and loosened the RAM timings, plus some small voltage bumps to Vcore and Vpll. Now I'm passing those 20K FFTs, go figure Edited by jamdox - 4/12/13 at 3:20pm