Stability testing

[Minusorange]

I've spent hours upon hours while running tests on my OC looking for definitive answers on what counts as a valid stability test and there seems to be no consensus.

There's people who swear by 24hour runs of Prime

Others will say Prime is to be avoided (especially for Haswells) and that tests like RealBench and X264 are better for showing stability

There's folk who say run Prime for an hour doing custom at 1344 FFT, others will say 10 runs in IBT on max will usually root out a dodgy OC

There's people who have said they could run Prime for 24 hours but crash running x264 and also the opposite.

So what is valid stability ?

I can run Prime on custom at 1344FFT, done 40 runs of IBT, ran x264 overnight without issue BUT I tried running the CPU:OCCT and it crashed after 10 minutes, however CPU:Linpack with AVX works flawlessly. Likewise Prime on custom with min 8FFT to max 4096FFT also yields a crash.

I'm on Haswell and I'm running close to my limit on temps, IBT, Prime, OCCT all push the limit to near 90 degrees so I'm wondering if it's more the temp causing instability than the actual voltage (even though Prime @ 1344 and IBT work fine) all the other tests don't really heat up my core and I hover anywhere from 60 to 70 depending on the test (currently running Asus Realbench bouncing from 68 to 70)

I know it probably comes down to personal opinion but if it can't run everything then surely it's not stable ? Or are we to not include the absolute extremes we find in Prime/IBT/OCCT simply because they are absolute extremes and nowhere close to realistic usage (unless you happen to enjoy using your pc to calculate extreme prime numbers) ?

 

[Silent_Scone]

You'll do well not to open a can of worms with that question.

Depends on the workload and the CPU architecture. Personally for CPU stability on newer platforms RealBench is sufficient assuming you're not entering arbitrary values hoping things work. IBT is a fairly weak test of stability by today's standards. The trouble is, and this is where arguments arise, not many people fully understand the implications of pushing OCCT and AVX versions of Prime for hours on modern CPUs. On the HEDT SKUs, you're talking 300 to 400w of current in some cases. It's a free country, but if you want to subject your CPU to that then there will be some degradation.

Memory stress testing comes down to two applications currently which are Google Stress App test, and HCI Memtest. Both good in their own right with the latter being better at catching errors with cache instability. AIDA64 is also good at finding cache instability.

Short answer, there isn't a magic bullet here, a pragmatic approach is needed depending on the platform - and simply use the machine beyond initial stressing.

 

[kully1972]

I just do a few runs of IBT. Realbench run and maybe prime for 20 mins.
Been overclocking for 15 yrs and TBH there's no substitute for real world testing.
Ive done comprehensive testing in the past and then it'll keel over In-game.
As an avid gamer I tend to play my games and no crashes after a week or two usually means it's all solid OC.
Battlefield 1. Rainbow six seige and Tom Clancy Wildlands are great testers, especially for memory issues.

 

[Minusorange]

You'll do well not to open a can of worms with that question.

Depends on the workload and the CPU architecture. Personally for CPU stability on newer platforms RealBench is sufficient assuming you're not entering arbitrary values hoping things work. IBT is a fairly weak test of stability by today's standards. The trouble is, and this is where arguments arise, not many people fully understand the implications of pushing OCCT and AVX versions of Prime for hours on modern CPUs. On the HEDT SKUs, you're talking 300 to 400w of current in some cases. It's a free country, but if you want to subject your CPU to that then there will be some degradation.

Memory stress testing comes down to two applications currently which are Google Stress App test, and HCI Memtest. Both good in their own right with the latter being better at catching errors with cache instability. AIDA64 is also good at finding cache instability.

Short answer, there isn't a magic bullet here, a pragmatic approach is needed depending on the platform - and simply use the machine beyond initial stressing.

This is the thing that gets me though I can run Prime95 on the majority settings without any problem, been running blend for the past 3 hours with no issue, yet if I do custom on 8k FFT to 4096k FFT for some reason it causes a crash at the second test, while Blend which basically does the same thing as that custom setting runs without problems.

So where do you draw the line on stability ? How does say @8pack do his stability tests for his super CPU's OCUK offers for sale there really should be some universal baseline amongst the community for what stable really means.

It's difficult for me to say my system is stable because that one single custom mode in Prime95 but if I ignore that my system IS stable, I guess it's like saying your house is water proof until the ground around it floods ? That one test being the flood while all the others are just like rain.

Thanks for mentioning Google Stress App too, not heard of it before so I'll be giving it a run in

 

[benjii]

In the functional sense, stability is subjective. Yes, total stability would mean no BSOD regardless of what you do. However, what if you never use AVX functions, why make it stable for that? If you don't know what AVX is or do know what it is and know you don't use it, forget about any stability test that uses AVX functions. Run Realbench or Overclock.net's x264 stress test v2 with a Unigine or 3Dmark stress test going in the background, overnight, and if its still running in the morning, consider it a job well done.

Just make sure, before you go to bed, that you watch the temps for 20+ mins to make sure that they are ok. You're fine to browse the net while you wait, the pages might just load a little slow.

 

[Silent_Scone]

The custom run is likely hitting the cache slightly harder. Like I say, there isn't a magic bullet here or a definitive answer. Anywhere where there is AVX involved, you have to acknowledge that if you want to overclock - concessions may need to be made. That's why Intel introduced the AVX offset.

 

[Minusorange]

The custom run is likely hitting the cache slightly harder. Like I say, there isn't a magic bullet here or a definitive answer. Anywhere where there is AVX involved, you have to acknowledge that if you want to overclock - concessions may need to be made. That's why Intel introduced the AVX offset.

Do you think it could be voltage for the vring then ? I'm running uncore at 35 and Vring at 1.1 (although my UD4H automatically boosts uncore to 40 to catch up with the 44 core mulitplier)

Same with CPU:OCCT it's not stable which could be same issue as that one Prime custom issue, while CPU:Linpack in OCCT with AVX works fine, I haven't been able to actually find any information on what the difference between the 2 tests in OCCT is though so I can't really troubleshoot as I don't understand what is being pushed more in the 2 tests.

Ideally I want to get the core stable before I attempt going 1:1 on uncore (doubt it will happen) but if that's what's causing the issue then I'll have to tweak and technically if I ignore that one Prime custom and OCCT it is stable and has cool temps (bar prime/ibt temps)

 

[opethdisciple]

The problem with any argument that says that overclocking is subjective and depends on he workload is that you can use any stress test you like but the PC wont fall over at stock settings unless it has defective hardware. This just tells you it falls over at all because of your custom settings.

If it falls over in one test but not in another, it simply was never stable in the first place.

So yes, if its 15k hours stable in Prime 95 but then falls over within 5 mins of OCCT, well then it ain't stable. Take it down a notch, get it OCCT stable, then go back to Prime and it should also be stable. As you've lowered clock speed or adjusted something down for stability.

If it's stable in one it should be stable in another. If it isn't it was never stable in the first place.

Stress tests are just pieces of software at the end of the day.

----

But yea OC is annoying, because so many people have so many different ideas on the subject.

I would never subject my PC to 24hours of 100% cpu load at high temps. I used to do 8hrs as I thought that was the right way to do it, but I've been told on here that 2-3 hours should be sufficient.

I think the magic is actually realizing when you have your cpu on the edge of stability, it might look stable but in reality it is literally dithering on the edge.

More annoying is when you CPU is stable for 10months, you upgrade to Win 10 and it crashes in game!

 

[new boy]

For me, playing GTA V for a few hours is the real stability test.

That game crashes overclocks that "stability test" programs are stable for hours on end

I do use realbench for dialing in a quick and dirty baseline OC to work with though.

 

[Silent_Scone]

Do you think it could be voltage for the vring then ? I'm running uncore at 35 and Vring at 1.1 (although my UD4H automatically boosts uncore to 40 to catch up with the 44 core mulitplier)

Same with CPU:OCCT it's not stable which could be same issue as that one Prime custom issue, while CPU:Linpack in OCCT with AVX works fine, I haven't been able to actually find any information on what the difference between the 2 tests in OCCT is though so I can't really troubleshoot as I don't understand what is being pushed more in the 2 tests.

Ideally I want to get the core stable before I attempt going 1:1 on uncore (doubt it will happen) but if that's what's causing the issue then I'll have to tweak and technically if I ignore that one Prime custom and OCCT it is stable and has cool temps (bar prime/ibt temps)

Possibly vring, but pushing this rail is what will degrade your CPU the quickest.... I've seen a handful of users do this with no more than 1.2v in Prime, sometimes less if it's run for long enough.

Just use the machine for what you need to use it for, or be another user that's pushed theirs till the gates are the width of a motorway.

 

[8 Pack]

I do none avx prime with 3d running at the same time to stress all controllers, Xtu with the same and real bench for a gaming pc.

Hci for memory.

For a workstation I will tailer testing to the company I am supplying them to. So if they are running avx etc. Avx 512 adds 15 degrees on temp and almost 100w on power draw on a 7900X so why subject the Silicon to this if it will never be used.

The best way in all instances is everyday use.

 

[Minusorange]

Possibly vring, but pushing this rail is what will degrade your CPU the quickest.... I've seen a handful of users do this with no more than 1.2v in Prime, sometimes less if it's run for long enough.

Just use the machine for what you need to use it for, or be another user that's pushed theirs till the gates are the width of a motorway.

I just wish I knew what the issue was that was causing the crash instead of being stuck with multiple choice lol

If it's temps I'm screwed because I can't push voltage higher for Prime
If it's voltage I'm screwed because I can't push the temps higher for Prime
If it's the Vring I "could" go 1.2v but like you say degradation becomes an issue so is it really worth it ? (I have got a good 4 years out of the chip already though)

It's interesting to see all the answers though as everybody has a different benchmark for stability

I do none avx prime with 3d running at the same time to stress all controllers, Xtu with the same and real bench for a gaming pc.

Hci for memory.

For a workstation I will tailer testing to the company I am supplying them to. So if they are running avx etc.

The best way in all instances is everyday use.

How long do you run your tests for ? 24 hours or less/more ?

 

[JasonM]

I've spent hours upon hours while running tests on my OC looking for definitive answers on what counts as a valid stability test and there seems to be no consensus.

Use the machine in it's normal day to day activity, if after a month the machine has not crashed it's probably ok.

Often a overclocked machine can crash at idle away from any CPU load.

However after overclocking you should run some limited load test, not so much to test if your stable but to make sure the CPU/Cooling are operating in safe temp limits.

 

[Adam123]

An hour or two AVX OCCT, 8 hours Realbench with full system RAM selected and some games. I'll sometimes do a Blend on Prime (non AVX) with Valley running in the back ground overnight just for good measure

For me if it passes that I'm fairly happy it will be stable

 

[DarkHorizon472]

For me, playing GTA V for a few hours is the real stability test.

That game crashes overclocks that "stability test" programs are stable for hours on end

I do use realbench for dialing in a quick and dirty baseline OC to work with though.

I do the same with BF1, realbench and other tests run fine but it will crash in under an hour on BF1.

Best to test what is relevant to how your use the machine unless you are overclocking as a test to compare to other people on similar systems.

 

[Silent_Scone]

Battlefield games have always been quite good at riddling out CPU instability.

And the best part is, you don't have to load up the CPU with 400w worth of current in a synthetic AVX bench with slightly more voltage just because it crashed in a game. Simply add 20 to 30mv vcore and you're there. This is the least taxing way to do things on a system that isn't data critical, and arguably if it is you shouldn't be overclocking anyway

 

[Minusorange]

When it comes to GPU stability what would you guys recommend ?

I know of Heaven, Valley and 3DMark but are there any other good bench/stress testing tools focusing around GPU or the full works of RAM,CPU,GPU ?