For the overclockers out there

[Dippndots]

Hey everyone, I was wondering for those of you running overclocks on your CPUs and RAM what sort of stress testing you guys do to make sure it's 100% solid.

I've OC'd a couple of Ryzens and Intels now for use in the office/rendering, and even with hours and hours of stability in AIDA64 or Prime 95, I routinely would get a crash here or there when rendering scenes in Corona. What's even more annoying is that these crashes can come after 10+ hours of rendering or even like the 3rd or 4th render. From there it's pretty just adjust the voltages until I stop getting the crashes and I'm officially stable.

It'd be handy to find a stresstest that can produce these instabilities before I fire up corona though. Any ideas?

[Juraj]

Prime95 is very configurable, but I am definitely not expert to know how to set up the instruction sets to make it most taxing. Right now, Corona is among the top but Corona doesn't utilize the CPU all the same during the full process and particularly denoising seems very taxing.
This would be good thing to establish.

[karnak]

I tried to document myself on this topic earlier this year.

I struggled a lot because of the amount of disinformation there is online (myths not based on actual facts, blind and aggressive devotion to a particular software or testing methodology and the list goes on..). It's hard to separate the bad information from the good one.

Here a list of the few things I have learned and tested:

• Keep open HWiNFO64 and monitor Windows Hardware Errors (WHEA) at the bottom of the sensors list, system might not crash even if there are WHEA errors listed, which is nevertheless a sign of an unstable overclock.
• AIDA64: System Stability Test (default) is useful for monitoring voltages and temperatures under plausible load
• AIDA64: System Stability Test (FPU only) is useful for monitoring voltages and temperatures under unrealistic continuous AVX load. Be careful with it because temperatures go up quickly.
• AIDA64: System Stability Test is not useful for testing stability, because it is unable to make a lot of errors show up.
• Prime95: before a certain version there were no AVX instructions used, recent versions have AVX instructions enabled by default, but you can disable them with a string option. Basically no need to use old magical Prime95 version.
• Prime95: Blend (default) test is very useful for testing stability, you can read in its logs every time there is an error. Sometimes the system don't crash even if there are fatal errors listed in the log, which is a sign of an unstable overclock.
• Prime95: Small FFTs (default) puts the system under unrealistic continuous load.
• Prime95: Small FFTs (with custom settings). From my understanding of it, most of the popular suggested values are based on quite old tests done by a user with electrical engineering background, but since many years passed and cpu architectures changed so much, those values are just magical settings. You know, like the ones that people put in render settings or material because someone 10 years ago told them to do so.

edit: fixed some sentences

[Dippndots]

Hey Karnak that's really good info. I'm embarrassed to say I'd never seen the WHEA section, and I'd never thought of checking the Prime 95 logs as I kinda thought, if it doesn't crash it must be stable! Haha.

Well, I'll be sure to revisit this in September when the 3950X comes out.

[karnak]

Happy to be of help!