Hacker News Reader

Comment by strainer | Hacker News Reader

When I used to tweak settings to the max i found that even heavy testing was no guarantee of stability in following months or years. With the system set on the edge of stability it can pass heavy testing but still be super susceptible to slow degradation of silicon and electronics and to transient conditions. I like to set things quick and cool running these days - i do quite quick stress testing to find roughly the best voltage/speeds and then add buffers of 5-10% voltage and sacrifice 5-20% speed to help guarantee some years of stability.

View on HN

That is why you back off from the bleeding edge after endurance testing it.

I have not done overclocking for quite a while, but general strategy was:

0. Get the right hardware - some hardware oveclocks better.

1. Depending on what you overclock come up with a short stress test. This would be GPU/CPU/RAM/PSU testing.

2. One component at a time. Identify top voltage/temp you are willing to tolerate. Overlock in steps of 5-10% nominal frequency. When the stress test fails increase the voltage by 5-10% until you reach the voltage limit. Once you get all of the data look at the frequency voltage curve and pick a voltage point (generally lower is better), back off a step or two from the frequency that passed and run an endurance test.

3. Run all components at the same time and stress test the entire system. The caveats here are to look out for power and temperature issues. I.E. make sure nothing get too hot (this includes the motherboard power regulators!) or voltage rails drop out of spec.

4. Slightly back off from the setup that passed 3 and you should be good to go for long term usage.

TL-DR: Overclock one component at a time come up with frequency/max temp/required voltage chart and don't try to run the hardware on the bleeding edge.

edit: list spacing...

I thought I was rock stable after lots of testing in one case, memtest, prime95, 'stability test' even thrashed the floppy drive. Later i wrote a java applet that hit the cpu cores in both regular and random millisecond patterns - it crashed the machine in seconds. Multicore load effects are something else to look out for then - take a few steps back from the edge and it all matters much less. Its worth it for a solid machine.

My takeaway from reading into these things was that at certain temperature breakpoints, lifetime decreases a LOT, way too much to want to get close to that.

So I tried to find the rough limits and wrote down idle/load temperatures for various combinations of settings/frequencies, and then picked my "sweet spot" at which temperatures started to increase a lot more per little increased performance. Granted, I did get a good case and CPU cooler, which I clean religiously, and of course there's always luck involved, but divided by the 8 years it's running this current computer is the cheapest I ever had, apart from having to replace the PSU because that was the one bit I paid less attention to, even the hard drives are still spinning, it's freaky. So the next one will have an even better case and cooling and then not get overclocked much or at all, either, that's for sure.

8 years is good going and Eco friendly :) When a machine is no longer the latest rocket, may as well run it cool too, except in winter if the heat is handy.

Reply on news.ycombinator.com