Why do people "torture" test their CPUs while overclocking?

beany_bot · 17 Dec 2015 at 11:06

Firstly, I don't claim to be any kind of hardware expert, I am very novice in this field. I fully understand testing you new cpu under heavy loads for a bit to see if it is stable, but I keep reading elsewhere online "Oh you must run Prime95 for a MINIMUM of 24 hours.

Coming from a non hardware background, I don't understand this.

I mean, if you upgraded the turbo on your car. You wouldn't sit a brick on the accelerator for 24 hours with the engine on the redline to torture/stress it would you? You would just take it for a few laps of the track to make sure everything is running OK?

Just trying to get my mechanically inclined brain round to hardware thinking.

Sin_Chase · 17 Dec 2015 at 11:32

There are many different schools of thought. The method one employs is largely down to preference and usage of system. A synthetic test might expose an instability that might rear it's head if you use your rig for rendering/computation tasks (which use certain extensions) By the same token it may not expose an instability that playing a game will. The idea behind 24/48/X Hour testing is that an instability may not present within a few minutes, or even hours. Once you are past the temperature equilibrium for all system components at that specific load going longer does not necessarily mean a more stable overclock. It does however demonstrate that the configuration can perform a heavy synthetic load for a long period of time, which serves to give a better confidence rating.

I personally will run a brief Intel Burn Test to check max voltages and get an idea of heavy load temps. If I am happy with both then I will just game on it. My system is a gaming system that sees rare computation/rendering tasks so it's a method which works for me.

You will find many people have a different term for 'stable' also. To some it's passing every test (synthetic, gaming or otherwise) for an extended period of time. For others it is being stable in whatever they use the system for.

Sods law is you get a stable under load system, but the second it goes into a C state the bugger BSODs/fails to wake

Frank Jeagar · 17 Dec 2015 at 13:17

This is a simple way of looking at it, but it isn't exactly perfect, it helps to understand it though I think:

Basically, unstable overclocks don't necessarily fail due to an accumulation of something (like temps or something) it can be seen as random.

Running a stress test maxes out your CPU in the hope of any instability resulting in a CPU failure.

However, if your CPU is "a bit unstable" the chance of it randomly failing is low, so you want to run a stress test for a really long period of time so that if it is unstable at all if will cause a failure alerting you to the instability.

Because even if it is only a bit unstable, one day you might be in the middle of something when it randomly fails which would be upsetting to say the least

lamboman · 17 Dec 2015 at 13:28

Really an apple and oranges comparison. Aftermarket parts for your car have been tested, likely with at least the bare minimum supporting modifications to ensure that the modification isn't going to break something. There's R&D behind the end result. If you randomly fit parts to your car free of any logic, that's similar to selecting a random voltage and frequency and expecting it to just work.

Also, you're testing for stability with varying loads, as opposed to load alone (if a motor can survive some torture, it's probably not going to blow at idle). An overclock can fail when you utilise a particular CPU feature, stress a particular system component or bus, etc (there are effectively multiple "types" of load).

This is why you don't want to run a single torture test to gauge stability. Realbench isn't bad however by itself but it is by no means conclusive.

In my case, I've tested with the latest versions of Prime95 to stress using AVX2 (and thus run a conservative clock and voltage, 4.1GHz at 1.1375V on a 5930K), and Realbench/games/Valley to complete the roundup.

As for the definition of stable, my personal view is strict and I can't see how anybody can argue otherwise - if the system can't handle stress under the same conditions as it would in a stock configuration, it isn't truly stable. A 4.5GHz, 1.3V overclock on an Haswell-E chip isn't truly stable when it can't run Prime95 without overheating and degradation.

Stable enough however, I have no problem with. For every other workload, if the above configuration ran without issue, it's stable enough for the user.

eddiew · 17 Dec 2015 at 14:03

In addition to all the above, I'd suggest that little bit of stress testing is checking that your power supply can deliver smooth, increased current for a sustained amount of time without any excessive voltage dips. In my mind, that's also a component that could behave worse as its temperature increases, so you definitely want to put a maximum-possible load through it for some time, to make sure that it will really drive the whole PC at full tilt.

lamboman said:
...An overclock can fail when you utilise a particular CPU feature, stress a particular system component or bus, etc ...

I was thinking about this; there's so many different sections on a modern CPU that it's hard to use all of them. Most tasks don't need hardware encryption acceleration, or video decoding, for example - but I wonder even whether sometimes there are chunks of CPU cache that just don't get used because there's not enough demand on the processor at that time.

wazza300 · 17 Dec 2015 at 14:08

Pre ivybridge/haswell you could stress the CPU with those programs but the latest Chips run so hot you can't use them anymore

A brief 5 run Intel burn test on maximum should be sufficient followed by game playing and general pc use to find any instability

Disco_P · 17 Dec 2015 at 14:21

Advantages:

Relatively quick to find issues
Very easy to use
Can be left for days without input
If there is an issue and everything crashes I lose nothing.

The last one is the big one for me. If I crash during a game I might lose loads of progress or an instability. might corrupt a game save. Same can apply if you work at your PC for example saving a large project file or exporting a video. A CPU going **** up could cost you a lot of money in lost time. The racecar analogy would be I'd rather it failed in the workshop where I can safely stop everything than on the track where it explodes turning me and the car into a fantastic fireball of death.

For me stable enough isn't a thing. Stable enough is unstable. My final stability tests are several consecutive overnight and day tests in P95, IBT, real bench then OCCT. These are about 18 hours each. Sometimes when I go away I'll just leave OCCT running for 100 hours.

I've not suffered any degradation having done this to my 2550k almost monthly for the last 4 years. That being said my temps have always been fairly low during these tests (around 70-80C) but voltage is high at 1.4v.

Rroff · 17 Dec 2015 at 14:30

There is very little benefit to a 24 hour torture test - but if stability is important then usually 2-3 hours of IBT and then bumping the voltage one notch if it passes in 99.999% of cases gets that. Stress tests aren't bullet proof as they tend to cycle synthetic loads that don't cover every possibility of real world use though they do tend to be fairly indicative.

Not a fan of just testing with the applications you normally use as there will likely be hidden instabilities that can bite you in the rear further down the line - i.e. I used to play lots of eve online for hours on end and finding the CPU was only "98%" stable in the middle of a fleet fight would end badly.

beany_bot · 17 Dec 2015 at 14:36

Disco_P said:
[*]Can be left for days without input

Without any degradation or damage or shortened lifespan to components?

EDIT: I see you have said no degradation.

new boy · 17 Dec 2015 at 14:43

A decent stress test saves time and effort.

Its a pain having your game crash when your playing with friends and having to add a notch of voltage only for it to crash again an hour later.

Anything that gets you closer to stability quickly is useful.

By a decent stress test, I dont mean 24h of prime though (not anymore anyway, I used to do that back in core 2 days)

lamboman · 17 Dec 2015 at 15:14

wazza300 said:
Pre ivybridge/haswell you could stress the CPU with those programs but the latest Chips run so hot you can't use them anymore

This is exactly the ambiguity that I was talking about when it comes to "truly stable" vs "stable enough", and it's why so much misinformation has been spread regarding P95/IBT/any other AVX2-capable benchmark. You can run these benchmarks if you're going for an as-stable-as-stock overclock, however your clocks will be reduced.

I know I'm being pedantic, however it's an important point to get across.

wazza300 said:
A brief 5 run Intel burn test on maximum should be sufficient followed by game playing and general pc use to find any instability

Not sure it is sufficient even for a "stable enough" clock. Huge, likely damaging current draw with tons of heat and throttling for a short period, it's not even close to any sort of real-world workload.

If big daily clocks are what you're going for, a four hour run of Realbench torture would be safer, more realistic, and is more of an all-encompassing test.

Truth be told, ultimately I agree with Disco_P's perspective however I'd say that the amount of testing (every month?!) is well within the realm of diminishing returns!

zipp0r · 17 Dec 2015 at 16:11

beany_bot said:
Firstly, I don't claim to be any kind of hardware expert, I am very novice in this field. I fully understand testing you new cpu under heavy loads for a bit to see if it is stable, but I keep reading elsewhere online "Oh you must run Prime95 for a MINIMUM of 24 hours.

Coming from a non hardware background, I don't understand this.

a cpu on the verge of stability can take hours to show an error, so to be 100% sure, we full load the cpu for a long period. 24hrs of p95 small ftt is a propper final test for the cpu overclock.

wazza300 · 17 Dec 2015 at 17:43

lamboman said:
This is exactly the ambiguity that I was talking about when it comes to "truly stable" vs "stable enough", and it's why so much misinformation has been spread regarding P95/IBT/any other AVX2-capable benchmark. You can run these benchmarks if you're going for an as-stable-as-stock overclock, however your clocks will be reduced.

I know I'm being pedantic, however it's an important point to get across.

Not sure it is sufficient even for a "stable enough" clock. Huge, likely damaging current draw with tons of heat and throttling for a short period, it's not even close to any sort of real-world workload.

If big daily clocks are what you're going for, a four hour run of Realbench torture would be safer, more realistic, and is more of an all-encompassing test.

Truth be told, ultimately I agree with Disco_P's perspective however I'd say that the amount of testing (every month?!) is well within the realm of diminishing returns!

you cant run those programs on modern cpu's it would throttle and cut out,ibt is a decent quick test for ivybridge and chips before that,then its upto the user to keep an eye out for any bsod codes and adjust voltages to suit,downlaod whocrashed or look in event viewer to see what caused the crash,a quick google usually tells you what the code relates too

but each to their own regarding testing,i don't see the point in hours and hours of stress testing,only to have it crash when in the idle state

setter · 17 Dec 2015 at 17:49

Agree with wazza. As someone who owned a 4770k and x2 4790k's, all would hit the 80's-90's at stock after 2 seconds in p95. Therefore, how can you reliably run such a program when overclocked using extra voltage which will only increase temps and therefore throttle.

lamboman · 17 Dec 2015 at 18:39

wazza300 said:
you cant run those programs on modern cpu's it would throttle and cut out,ibt is a decent quick test for ivybridge and chips before that,then its upto the user to keep an eye out for any bsod codes and adjust voltages to suit,downlaod whocrashed or look in event viewer to see what caused the crash,a quick google usually tells you what the code relates too

So you're telling me that Intel have a new feature (that will primarily be adopted in the HPC/Big Data space, huge markets for them) that their processors can't run? There's a huge difference between stating that they can't run those programs (false, they're optimised exactly for those processors!) and that they will struggle to run them when heavily overclocked (which is true for big clocks). IBT works well for IB and earlier because they don't feature AVX2.

wazza300 said:
but each to their own regarding testing,i don't see the point in hours and hours of stress testing,only to have it crash when in the idle state

Agreed. As has already been mentioned, a mix of tests is what you want.

setter said:
Agree with wazza. As someone who owned a 4770k and x2 4790k's, all would hit the 80's-90's at stock after 2 seconds in p95. Therefore, how can you reliably run such a program when overclocked using extra voltage which will only increase temps and therefore throttle.

I've got no experience with benching those chips (I'm running Haswell-E), my guess would be poor heat dissipation due to the non-soldered IHS. P95 for me at stock peaks at around 60-65C IIRC, and at 4.1GHz at around 80C (NH-D15).

setter · 17 Dec 2015 at 19:06

lamboman said:
I've got no experience with benching those chips (I'm running Haswell-E), my guess would be poor heat dissipation due to the non-soldered IHS. P95 for me at stock peaks at around 60-65C IIRC, and at 4.1GHz at around 80C (NH-D15).

Yep, the non soldered ihs and probably the addition of the vrm's on haswell/devils canyon led to much higher load temps when running any AVX instruction stress tests. Devils canyon was supposed to feature a new and improved thermal paste according to intel. But that turned out to be false. Im running a 5820k haswell E now at 4.5ghz on 1.300v, it is a cooler chip in general compared to the 4790k's i had, stressing with XTU i do have one core that hits 83c though, the rest are 76-79c. But i suspect that could be down to poorly applied thermal paste. When gaming the hottest core gets to 64c, 11c lower than my last 4790k which was running at 4.7ghz on 1.300v.

Disco_P · 17 Dec 2015 at 23:39

I'm so glad I invested in sandy bridge. None of this faff with cruddy tim or synthetic benchmarks causing excess heat. Even the overclocking seems simpler, looking at new guides I see ring voltage, cache clock, uncore etc.

Best £150 I've ever spent (equal to an ssd)

lamboman said:
Truth be told, ultimately I agree with Disco_P's perspective however I'd say that the amount of testing (every month?!) is well within the realm of diminishing returns!

Every month is far from necessary. I just can't stop messing around with my chip, trying to get offset voltage working, chasing a stable 5Ghz (1.51v anyone), attempting to lower volts at my 24/7 with vccio or pll.

You don't need an endless supply of £800 CPUs, pots of LN2 or even super high end mobos to have fun with overclocking.

Jumper118 · 18 Dec 2015 at 06:57

I never stress test any more. I overcloxk so many CPUs that I can tell if its going to be stable or not. Even if it isn't I don't really care. I know most bsod of codes by now so I know exactly what to do if it does blue screen.

Disco_P · 18 Dec 2015 at 08:07

Jumper118 said:
I never stress test any more. I overcloxk so many CPUs that I can tell if its going to be stable or not. Even if it isn't I don't really care.

Can you elaborate? I've got images of you studying the starting windows animation to look for instability or whispering to the motherboard.

lamboman · 18 Dec 2015 at 11:15

Disco_P said:
Can you elaborate? I've got images of you studying the starting windows animation to look for instability or whispering to the motherboard.

"The Clock Whisperer".