Best way to stability check a CPU OC

[Rokha88]

Hi everyone,

Since a few days, I am trying to get a slight OC (4,7Ghz@1.225v BIOS) on my I7-7700K, but as it is a hot chip, I am quite confused by the very disparate results I get from a stress test to another, especially about temperatures.

What software would you recommend to stress-test the stability and for how long ? I have read that Realbench is a good test as it is quite realistic. Some recommend to run the benchmark for a few hours, some recommend the Stress-test for a few hours ? What about it ?

Another concern is that on IntelBurnTest, Prime95 (most recent version) & OCCT Linpack, I get INSTANTLY crazy temperatures (90 C°/194F° average), how can it be so hot instantly ? I have read that it may be due to the AVX instructions, as I don't get these temps on other software, or on older versions (example V266 of prime95 it seems to work great). The avg is then 73C° (163F°) Should I bother with it ?

For information, I have Maximus VIII Ranger, I7-7700K (4,7GHZ@1.225v bios), Scythe Fuma REV.B (X2 fans) CPU cooler, and Zalman Z11+ Case (1 front intake / 1 top exhaust, 1 rear exhaust).


Thanks for your answers :-).

Bye.

[Falkentyne]

RedSector73 wrote:
Aida64 - System Stability Test
Can stress
CPU
FPU
Cache
System memory
Local disk(s)
GPU

As a whole or individually

I'd call that comprehensive stress test, besides once got stable in it never found anything that prove otherwise.

As Bass junkie said, AIDA64 puts a rather balanced load on the CPU. It isn't a bad tester but you have to rely on more than one test.
I came up with minecraft "Spam" loading myself.
My testing found that I can pass AIDA64 Stress FPU (This is an AVX test too) at a lower bios set voltage, than minecraft! Yes, minecraft requires a higher voltage.

The reason for this are transients. Minecraft puts an erratic up to 100% full core load on several different cores at once and gets shuffled around. This causes transient response to plummet down lower than what you get in something like Cinebench R20 or AIDA tests.

Another good test is Battlefield 5.. Loading certain maps can put a 100% violent transient core load on all cores at once, even all 20 threads. The ripple means that you can get a CPU Cache L0 error when this happens, while nothing happens at all AIDA tests, then you need more voltage. That's just how it is.

That's why I come up with my own tests. Super heavy load stress tests like Prime95 or AIDA stress FPU are great for finding your absolute minimum voltage you need to be stable, because the transients are stable and small because the load doesn't change violently during the test. Transients can't be picked up on voltage sensors. You need an oscilloscope. That's why those transient heavy tests I listed can be better as a transient voltage dip can go below the sensor "Vmin"--and then you're unstable.

If you really want to see some nasty transients, run Prime95 128K AVX in-place fixed. Buildzoid has an oscilloscope video of that test. The transients literally cause a MUCH greater "peak to peak" voltage ripple when you run that, while the RMS and sensor voltages only show the mean voltage. But it's the minimum transient voltage drop down point that determine how stable you really are!

[bass_junkie_xl]

adia 64 is walk in the park for a cpu pass adia 64 then bsod / whea error on mine craft .

use asus real bench
occt
blender benchmark
x256

[HisEvilness]

You can run prime95 with custom large FFTs to emulate a gaming load, normal Prime95 settings are OP but offer a good way to really push an OC to the brink so you know for sure it is stable. Add CpuSupportsAVX=0 to local.txt to turn off AVX.

[menaquinone]

I have found that Prime95 (with the below settings, seen in screenshot) is the fastest to error (compared to OCCT, Intel XTU's stress tests, powermax, cinebench) and it causes the highest temperature(yes, it is using AVX and AVX2 even). Testing with others can cause you to waste minutes or hours until you see any error(s), if you even see any at all! That's why I always use Prime95 first, it usually errors in under 1 minute, then I know to decrease by like 0.010V or at most 0.02V and it's usually stable there. The longer it takes to error, the less I've to decrease in voltage. Only afterwards I can test with others like OCCT, but it would be a waste of electricity because they won't crash it...

That being said, my testing experience is limited, so feel free to consider other people's advice instead.