It dawned on me that there's an unchallenged assumption in standard advice for overclock stability testing: We declare an overclock stable if it survives an uninterrupted full load on all cores for 24 hours.
My current stress test is a representative parallel computation of the sort I build my machines for, that takes 45 minutes on a 4.5 GHz 2600K. (By giving up that last 10% I get a nearly silent machine on air.) My script loops with a five second rest between 45 minute iterations.
Just as most plane crashes are within sight of the departure or arrival runway, all freezes and random reboots as I play with BIOS parameters are either at the start or the end of one of these 45 minute sprints.
Pretty obvious once one asks the question, right?
This reminds me of the load line calibration debate. I'm now an agnostic; the Official ASUS P8P67 Series Overclocking Guide and Information recommends enabling LLC for overclocking with my ASUS P8P67 Pro motherboard, based on their empirical results. On the other hand, I still fondly recall Kris Boughton's writings here at Anandtech on the idiocy of using LLC.
In lay terms the board voltages bounce around like crazy, going back and forth between full load and idle. LLC messes with this, for better and worse, so we each decide for ourselves if LLC helps.
So who are we kidding, calling an overclock stable if it survives a continuous full load with no idle rests? Hammering back and forth between full on and idle is both more realistic, and more challenging.
More to the point, I crash quicker when I'm going to crash, which sure speeds up testing.
My current stress test is a representative parallel computation of the sort I build my machines for, that takes 45 minutes on a 4.5 GHz 2600K. (By giving up that last 10% I get a nearly silent machine on air.) My script loops with a five second rest between 45 minute iterations.
Just as most plane crashes are within sight of the departure or arrival runway, all freezes and random reboots as I play with BIOS parameters are either at the start or the end of one of these 45 minute sprints.
Pretty obvious once one asks the question, right?
This reminds me of the load line calibration debate. I'm now an agnostic; the Official ASUS P8P67 Series Overclocking Guide and Information recommends enabling LLC for overclocking with my ASUS P8P67 Pro motherboard, based on their empirical results. On the other hand, I still fondly recall Kris Boughton's writings here at Anandtech on the idiocy of using LLC.
In lay terms the board voltages bounce around like crazy, going back and forth between full load and idle. LLC messes with this, for better and worse, so we each decide for ourselves if LLC helps.
So who are we kidding, calling an overclock stable if it survives a continuous full load with no idle rests? Hammering back and forth between full on and idle is both more realistic, and more challenging.
More to the point, I crash quicker when I'm going to crash, which sure speeds up testing.
Facebook
Twitter