What is stable system? Is Prime the best method? - Post your thoughts here

[Pedro Rocha]

In my mini-revew of DFI 680i, virtualrain post a question that for many of you is recurring: what is the best method to test system stability?

All start with my toughts

If you guys insist I can try some quad cores primes, BUT (and that's a big but) I do not understand why some of you guys are so blinded with Prime.

Is there any real life application / game / benchmark that put some much stress on the components as prime?? I don't think so

I do overclocking for near a decade and I always follow the 10 - 15% rule for 24/7 use and I never had stability problems.

For instance if I had a CPU that run most of the benchs (superpi / 3mark, etc) at 4.000Mhz, for daily usage I always run it wih less 10 to 15% of the max OC, more or less 3.500Mhz in this case.

Please do not flame this, it is just my opinion. I respect others opinions, if you consider a system stable only if passes a 100% stress prime it's ok - use it and stress it..

The most important thing in OC it is not to have stress (prime or other) but have fun a enjoy your system

Here is what virtualrain

Not blinded by PRIME, just "grew-up" that way

It seems overclockers are generally getting more slack with what is considered a good overclock. While you have a good argument, there's also something to be said that if your system can run PRIME stable at stock it should be able to when overclocked as well... otherwise it's not a good overclock. Unfortunately, following such doctrine results in much less attractive thread titles.

I think if you scale back 10-15% you are probably well within the margin of stability... so why not test? My board boots at 3.8GHz but can't do much without crashing. At 3.6 it's PRIME stable for 8+hrs. That's only backing off about 6%.

Ultimately though, I respect your opinion as well and each person has to be satisfied with their own overclock. It's just getting harder and harder to compare overclocks now since everyone's standard for what is stable varies.

He raise a interesting question: what is the best method to compare stability in overclocked systems?

Some defend Prime and their variants, like Others.

I dont have a firm opinion, since I have 2 diferent rigs one to bench (play my fav game: 3Dmark2001 ) and other for 24/7 that is almost noiseless and run with a very moderate OC.

The idea of this thread is to post your toughts about this question.

Any post related to that discussion will moved to this thread.

This way we all can contribute to have XS more clean and easy to read, by avoiding this interesting discussion to be spred in many threads.

Thank you

[Pedro Rocha]

Fhpchris opinion:

The 10%/15% rule seems a little optimistic, if you have an x6800 that does ~4000-4100 benching everything, then why run it at the same voltage at 3500?

If the board does not like to pass prime blend, then that has to be said; this board would not be the first one to do so -- the AW9D max was a great board, but would not pass prime blend above 400 FSB to save its life. Was the system still stable? Yes. small FFT would pass for hours...

To run a 3.9ghz+ daily rig on water, you will need a tool like prime, even if its not the best solution. Now, personally, I only run the quad @ ~3300ish on stock voltages daily, and when I had an X6800 I booted @ 6*435 and didn't notice for normal daily tasks (not gaming/encoding/blah clearly). I guess I agree with you, but I still want to keep using prime as a good tool for finding absolute stability; you said it yourself:


Prime stresses a system harder than most anything really will, and thats the point

[Pedro Rocha]

Raju opinion

I use a lot of occt now for stressing work, seems to hit home hard in a shorter duration, became tired of sitting thru hours of orthos, not against either really, though if I can get to the point faster it helps...

regards
Raja

[cadaveca]

Hi Pedro...

I firmly beleive in a set system of stability testing using a variety of applications, in a specific order, to ensure 24/7 system stabilty.

You'll find that I almost never post "top" 3Dmark numbers...this is because before I ever run a bench, I must consider the system stable enough that any task can be performed on the system, memory, cpu, or video intensive, and even all three.

There are applications for each type of testing, and errors given by tests can tell you what the problem is in your overclock, however, delving into such is not something that can be undertaken very easily...

I highly advocate using Prime95, one instance for each core, starting @ 8k test, to largest FFT that can fit within cpu's cache, for cpu testing. For complete pass, system must be able to complete 4 full runs of each fft size, sometimes taking many days.

While doing such testing ,I monitor system power consumption, temps, voltages, droop, and a host of other things, each to ensure things are working right. I cannot divulge what it is that I watch, as it really depends on what it is that I am trying to push....FSB testing take a different approach than max MHZ testing.

I DO NOT advocate ONLY testing with prime95...I like to have some music playing foer the entire test(muted if i do not care to hear it), as well as some webbrowsing and such. This allows for a varied load to be put on the cpu while running prime, and the "drops" in cpu usage can trigger faults caused by poor power or the like. I also like high network traffic, again, testing board power delivery and chipset stability(chipset ina partial way...relative to the load put on it by cpu).

memory testing is Memtestx86+, and I generally find 4-8mhz less will be fully stable in windows XP than memtest, and 5-12mhz less for windowsVista. Personal thing tho...

Chipset testing varies on platform. AMD has memory controll in cpu, Intel does not, so stability testing such can be problematic, but I always find it's best to test by varing voltages to find limits of stability, and then tweaking within the presented range to find what's really stable(always tending to the lowest possible voltage, of course.)


I am fortunate enoguh to always have a host of parts to play with, so when it comes to testing, I have "known stable" parts that allow me to then test each new component extensively. Things like chipset straps, memory alpha timings, etc, all play a role in how all parts perform, so simply testing a part in one platform is not enough to truly find "the max" of a given component, as I sure you will agree.

That being said, you really need proper facilities in order to do proper testing, and proper protocol to how that testing is run, or it's not proper testing. Again, a given.

I run a D&R business(Develop and Re-use), so my testing is pretty strict...i need to be able to integrate components in a variety of working environments, each of which can place unseen demands on components in use by regular users, so i test things alot. Our motto is "Life...Better"...and running unstable systems doesn't make life better.

[Raja@ASUS]

tbh stability testing is only required for some in the public domain, with experience you can pretty much dial in a system overclock in a matter of hours. Best as in all things that become sectarian, is to adopt a 'live and let live strategy'..

One can often find results either way, forcing everyone to adopt the same stance is not required, as there will always be enough of each. As with all things, once something has a mass following (bench vs stable type reviews), it is best not to read if you don't like it, attempting to force a user to change a review type philosophy represents a certain amount of naivety.

regards
raja

[uOpt]

The point about mprime is not that it puts so much more stress on the system.

The point about mprime is that it detects most errors due to internal checks of intermediate results.

Folding at home, for example, is not much less stress, but the reason why you see less errors is that errors that appear are not reported. You will only notice if the error damages anything vital to running the program, or if it messes up your OS. But any wrong computational results you will never notice.

[7499richard]

I use Prime95 to test system stability, and I consider my computer to be rock stable if it passes 12+ hours of blend test on prime95.

What I usually do when overclocking is this,

If I change ram settings, I will run memtest86+ before I even attempt to get into windows. I run this for about 2-3 hours, if its stable then great. I then boot to windows and run the blend test for 12+ hours, if it passes then great.

But I don't believe that the temp of your cpu after 12 hours of blend test is the hottest it will get. So if it passes all that I then go for a max cpu temp, by running Prime95 small fft's, as this get my cpu the hottest. I run that for at least 6 hours, and I consider those temps to be the max I will ever experience with my current oc.

Thats what I use, and its proven to be very helpful to me, and it works great.

Now if only there was an awesome way to test GPU overclocks for maximum heat and stability?

Thanks

[railmeat]

"For instance if I had a CPU that run most of the benchs (superpi / 3mark, etc) at 4.000Mhz, for daily usage I always run it wih less 10 to 15% of the max OC, more or less 3.500Mhz in this case"


good point,and i agree %100

[lowfat]

I am a firm believer in Orthos. MemTest and SuperPi would be 8 hour and 32M stable, and Orthos could still find an error using blend within an hour. For myself I like to have my systems 8hr small fft and 24hr blend stable before I'll use it for a 24/7 setup.

[Raja@ASUS]

that's fine as long as you are certain that the prime code is fully compatible with every cpu and chipset, take a look at mersenne where the code has been sped up or slowed down to make it compatible with certain setups, I kinda lost faith in pushing new hardware with orthos/prime until I can be absolutley positive the software code is fully sound.

[7499richard]

raju,

what do you use then?

[Raja@ASUS]

a variety of things, cross comparing with OCCT, 3d tests and plenty of gaming.
Still run some prime, but if I find it tripping up too much I see where the other tests fall at the same clock. Some of the strategy based games like C&C 3 will weed out a weak overclock pretty fast.

regards
Raja

[kiwi]

My opinion

1. Always start with OCCT 30min
If it passes then there is a very high chance that your system will be totally stable with any other application

2. a. A few loops of 3dmark (I prefer 06 with cpu tests) or better solution b
b. Gaming a few hours, I am playing BF2, UT2k4 was good for single core

3. Run any BOINC project 24/7, D2OL, folding etc. I am running WCG (it also has error checking). These projects stress cpu about 5-10C lower per core than orthos but that's fine really, not many applications run as hot as orthos 24/7




You might also want to find a few stable oc spots
For example, my current cpu is as follow
3.9 1.55V
3.8 1.5V
3.6 1.4V

In winter I run 3.9 but now ambient temps are 10C higher so I backup to 3.6, dont really need to heat it now

[virtualrain]

The point about mprime is not that it puts so much more stress on the system.

The point about mprime is that it detects most errors due to internal checks of intermediate results.

Folding at home, for example, is not much less stress, but the reason why you see less errors is that errors that appear are not reported. You will only notice if the error damages anything vital to running the program, or if it messes up your OS. But any wrong computational results you will never notice.

QFT! (This is one of the best posts in this thread!)

If you don't test your system with a program that detects computational errors, why test? Using a game or benchmark that could be throwing all kinds of errors that are ignored doesn't make any sense to me. If you can run such an app for any period of time on an OC'd system all it means is that you got lucky, not that your OC is stable.

Using a game seems perplexing to me. How can one question the integrity of code in PRIME but then use a game for stress testing? Game crashes are far more likely to be the result of driver or game bugs in my experience.

I guess there are three kinds of testing one can run, and it's important to distinguish between them:

1. Stress testing. IMHO, this is where you load the system with as much heat as possible. TAT is probably the best such tool for Intel processors. I trust Intel to have done their job in binning my product not to have to repeat this kind of testing.

2. Stability testing. IMHO, this is where you run continuous computations that are double-checked for error and ANY error halts the process. OCCT, PRIME, Orthos, etc. are all good stability testing tools. Time here is what is hotly debated. 30min. is not enough in my opinion. If your computer can run a stability test for 8+hrs at stock settings, why shouldn't your overclock? If you computer can't run a stability test for 8+hrs at stock settings, then you probably have bad products or then you may start to question the integrity of the code.

3. Application testing. IMHO, this is where games, benchmarks, folding, etc. come in. All this will tell you is that your selected game, benchmark, etc. runs (at least at the moment... an unstable system could cause a fatal error anytime).

I personally feel that the overclocking community is deteriorating because the commercial pressures on betatesters, reviewers and journalists to show wild overclocks is causing them to cast aside principles that defined overclocking just a few years ago. A screenshot of SuperPI32M completing seems to be all it takes now to legitimize an overclock... am I the only one who thinks this is a bad thing?

[thomas66]

my feeling is that stability testing specially in the overclocking comunity got too much attention in the recent years.

when you overclock and you cant provide 12 hours ortos dual prime or dual super pi 32m - than you are a noob.

a stable systems for me is - when i can use it 24/7 without any crash related to overclocking.

so far i always went fine with a super pi 32m/1 hour orthos run. never did more for stability testing - i only went further when i had crashes in 24/7 use of cause than i would do more investigations but that happened only once so far for me.

[NLight95]

QFT! (This is one of the best posts in this thread!)

If you don't test your system with a program that detects computational errors, why test? Using a game or benchmark that could be throwing all kinds of errors that are ignored doesn't make any sense to me. If you can run such an app for any period of time on an OC'd system all it means is that you got lucky, not that your OC is stable.

I personally feel that the overclocking community is deteriorating because the commercial pressures on betatesters, reviewers and journalists to show wild overclocks is causing them to cast aside principles that defined overclocking just a few years ago. A screenshot of SuperPI32M completing seems to be all it takes now to legitimize an overclock... am I the only one who thinks this is a bad thing?

I've made a similar argument as yours in several forums, VR. Web sites like Anandtech confirm stability with only a 30min Orthos Blend. I'm sure other sites have done the same. There's immense pressure to throw out the high numbers to attract readers that ethical journalism appears to be cast aside. It does a great disservice to readers when they go to lay down their hard-earned cash for a component only to find out later it doesn't stably perform anywhere close to what was in those "great" reviews.

SuperPI is a performance utility and nothing more. It's useful for comparative analysis but not as any indicator of stability. It is no more an indicator of stability than one run of 3DMark06, which seems to take about the same length of time.

I believe Prime/Orthos is perhaps the most reliable method so far to determine long-term stability. But as others have noted it is extremely time consuming, not to mention all the electricity being wasted. I believe we really need a tool that will yield results much quicker.

My brother (a DBA) runs an Oracle script that will crash his system within 15-20min despite being able to pass a 12+hr Orthos run, so even Orthos is not the end-all be-all of stability. Then again, very very few hobbyists ever see that type of environment. I wish BOINC had an error reporting mechanism--at least that way the computing cycles are being put to good use.

Right now I'm fiddling with Everest Ultimate's stability test. I really don't know how thorough the test is at determining overall stability (no GPU test) but it really loads my system down during operation. I've used OCCT but don't like that it relies on third party software (MB Mon, SpeedFan, Everest) for monitoring system functions.

I've considered a somewhat unorthodox alternative: Dropping the CPU fan speed in half or a third and running the same test that passed previously. This would perhaps account for higher ambient room temps during different times of the year. You never know when the AC will go off or fail. But here again, more time & resources being munched.

[Pedro Rocha]

Good to see a good, informative and open discussion here - let's keep your motives coming

Going sticky

[uOpt]

SuperPI is a performance utility and nothing more. It's useful for comparative analysis but not as any indicator of stability. It is no more an indicator of stability than one run of 3DMark06, which seems to take about the same length of time.

Actually SuperPi is excellent as a stability test because it has about the same amount of internal checking on intermediate results as mprime has.

It's not as suitable for CPU clock testing as mprime is because mprime heats up the CPU a lot and superpi doesn't.

SuperPi is excellent in discovering memory that runs at too high speed or too low timings, and mprime doesn't test that at all.

As a result, all of SuperPi, memtest86 and mprime are all strictly required, missing one of them isn't testing at all.

[Raja@ASUS]

prime has no ability to stress the pcie load of a motherboard, I can think of instances where prime ran for hours but 30 seconds of 3d would fail. Bottom line there is no glorious all in one test.

Prime lovers always use the rationale of well if it runs @ stock it should run when oc'ed, the same can be said for many well coded games. Believing so heavily in one element always has defects as in all things. Irrespective of claimed experience, such belief systems are noobish at best, Uopt actually has the mental flexibility to understand the need for a variety of tests..

Live and let live, when it comes to a mans stability, if you don't like it, look elsewhere for your kicks..


The world will never dry up of sectarian belief systems, there are always enough of your 'beloved' to satisfy your curiosity..lol..


regards
raja

[NLight95]

For relative graphics (DX9) performance I find Aquamark3 to be quite useful. It's free, runs quick and gives consistent results. Real-Time HDR is also good for heating up your vid card and checking GPU core & mem temps. ATiTool is helpful in testing for artifacts, even on Nvidia's cards. I use CineBench and GLExcess for OGL testing, as well as Doom 3.

Ntune has a stability test but it doesn't run properly on my system--it always stops with one minute left. Sandra has a system (burn-in) test but I just can't bring myself to trust it. I'm starting to like Everest Ult's stability test. It's $30 for the software so I'll have to decide whether it's worth it in the long run (using the trial edition now).

I guess what it comes down to is that we all have our own way/methods of determining stability. That's fine fine for personal use but not for an OCDB. An OCDB must adhere to some standard suite of tests--whatever they may be--it's the only way all entries in the db are on an even keel. Of course, then everyone argues about what should be the standard.

[Raja@ASUS]

I usually run aq3 at 4.8 ghz or so, with timings so tight that the machine can trip at an minute, it's quite easy to pass...

[NLight95]

I usually run aq3 at 4.8 ghz or so, with timings so tight that the machine can trip at an minute, it's quite easy to pass...

Remember I said AM3 is useful for relative graphics performance. I never said it was an indicator of stability. I use AM3 to see how much a GPU oc has netted me. I use ATiTools to check for artifacts and let RT HDR run overnight for a final check.

[Raja@ASUS]

As the threads primary concern was/is full overall/simultaneous stability, my initial statement was about overall board stability and how Prime running stable for lengthy periods, does not necessarily mean that all 3d will run.. Even a GPU test will not confirm how well the NB will handle full data through put from all areas at the same time.

The main gist of this sideshow bob slot, was how reviewers don't want to run big hours of prime runs to keep Joe Public happy. Times change, as does software, this stable/benchmark lark, is always gonna remain a personal thing for everybody, reviews always feature around the stronger elements of a board rather than the weaker, sometimes we find exceptions. Question is who's prepared to wait and let the dust settle before hitting the buy button these days? Certainly most who spend large amounts of time, dropping woots and banana smilies, not to mention the 'thanks for sharing the lab pics, you are so kind to feed us this marketing material', at the mere mention of 500fsb are not able to wait for the blister to pop on their buy button finger. Don't wanna fall prey to hype? Cure your affliction of 'PC overclocking without rationale' addiction first (none of us seem able), then perhaps later things will slow down a little giving both parties time to make more rational decisions and to release well tested hardware.

If I were submitting to an OCDB, no doubt I would follow the time requirements, however after much experimentation I found running 30mins to an hour of OCCT was absolutely fine (for me personally), anything that managed to pass the 30mins test generally does not trip up elsewhere on my system. Prime seems to be structured differently and does not cycle the default blend element of the fft sizea around as quickly, needing longer runs to 'hit' various areas.

Even with all this in hand I am no longer naive enough to believe in any one tool for stability, there always the possibility that the board AND software may be at fault. A good example of the was the AWD9 Max, not great at Orthos blend, but it could run anything thrown at it way higher than the failure point, who knows why?

Both sides of review strategy can create false expectations, the idea is to run a little of everything to give just a small insight, not to give outright guarantees, things are just too variable, that's why for now I will stick with 30 mins to 1 hour of OCCT, it's enough for the way I run my system, until I get the feeling that something is no longer valid with this type of testing, any doubts and I move on. The OCCT runs I make are always lower down than the benchmark numbers, 3.2ghz to 3.6 ghz, which is way more realistic for 24-7 use, depending on cooling of course..

The rest? All one can do is state that they played a few hours of gaming and used their pc in day to day tasks without issue...

regards
Raja

[WiCKeD]

This thread could have been ended after cadaveca's illustrative post and I applaud a 3DTeam member for making this statement. There have been some other solid points made, but I think he captured the essence of this problem.

One can often find results either way, forcing everyone to adopt the same stance is not required, as there will always be enough of each. As with all things, once something has a mass following (bench vs stable type reviews), it is best not to read if you don't like it, attempting to force a user to change a review type philosophy represents a certain amount of naivety.

I think you are undermining what you are saying in later posts. Yes, there are threads where people give you an accurate picture of their overclock, but they are rare. When you collect results and see X overclock 1000 times with a certain processor/memory, it mentally muddies away any realistic expectations. With the right intentions, respected people have the power to change the behavior of others and shift standards. This is a moderated internet forum with the power to make these kind of changes.

While Prime/ORTHOS is not a perfect measure of stability, that is not a justification to AVOID running it! I see this excuse all the time. As was said, it is a reason to use ORTHOS in addition to other benchmarks. Are there real life apps that stress as much as ORTHOS - yes. Prime95 is one of these programs! That was its original purpose. Just because you do not use them, does not mean others will not. "Gaming stable" may work for you, but when you are reporting results on the internet, you are affecting others' hopes and purchases. Posting unstable results provides artificial expectations and damages the OC community as a whole. Put aside your own selfish ambitions for a second and think about how you are affecting others.

If I had to pick a definitive benchmark that can be run in a reasonable amount of time and that stresses the system substantially it would be OCCT. I would be extremely happy if everyone were posting a screenshot of this with their "OMG LOOK AT THIS L33T0RZ OVERCLOCK!!!" Then we will see just how much overclocks change...

Still, ORTHOS is possibly the best utility for quickly testing stability and nailing down your approximate overclock.

Another issue is how quickly people are ready to embrace manufacturer's own benchmarks. Hold them to the same stability standard as everyone else, so they are forced to provide more realistic results. Threads like these should be exclusive to their own sections or include some type of disclaimer. They have access to hoards of different parts where they can cherry pick results - people need to be hit over the head with this and understand their motivation.

This subject has come up many times on XS and every time it is ignored. XS is not the only forum guilty of this, but they need to be setting a better example that others can follow. That can be changed with its leadership and I am glad to see this being looked at again seriously.

[Raja@ASUS]

This thread could have been ended after cadaveca's illustrative post and I applaud a 3DTeam member for making this statement. There have been some other solid points made, but I think he captured the essence of this problem.

I think you are undermining what you are saying in later posts. Yes, there are threads where people give you an accurate picture of their overclock, but they are rare. When you collect results and see X overclock 1000 times with a certain processor/memory, it mentally muddies away any realistic expectations. With the right intentions, respected people have the power to change the behavior of others and shift standards. This is a moderated internet forum with the power to make these kind of changes.

While Prime/ORTHOS is not a perfect measure of stability, that is not a justification to AVOID running it! I see this excuse all the time. As was said, it is a reason to use ORTHOS in addition to other benchmarks. Are there real life apps that stress as much as ORTHOS - yes. Prime95 is one of these programs! That was its original purpose. Just because you do not use them, does not mean others will not. "Gaming stable" may work for you, but when you are reporting results on the internet, you are affecting others' hopes and purchases. Posting unstable results provides artificial expectations and damages the OC community as a whole. Put aside your own selfish ambitions for a second and think about how you are affecting others.

If I had to pick a definitive benchmark that can be run in a reasonable amount of time and that stresses the system substantially it would be OCCT. I would be extremely happy if everyone were posting a screenshot of this with their "OMG LOOK AT THIS L33T0RZ OVERCLOCK!!!" Then we will see just how much overclocks change...

Still, ORTHOS is possibly the best utility for quickly testing stability and nailing down your approximate overclock.

Another issue is how quickly people are ready to embrace manufacturer's own benchmarks. Hold them to the same stability standard as everyone else, so they are forced to provide more realistic results. Threads like these should be exclusive to their own sections or include some type of disclaimer. They have access to hoards of different parts where they can cherry pick results - people need to be hit over the head with this and understand their motivation.

This subject has come up many times on XS and every time it is ignored. XS is not the only forum guilty of this, but they need to be setting a better example that others can follow. That can be changed with its leadership and I am glad to see this being looked at again seriously.

last post from me, I already stated anything I do has either benchmark or some OCCT work (when I find variance I drop them and move on), what I stand by however is that I follow neither blindly or like sheep.


When the coders can come forwards and categorically state they are checking their code for compatibilty every few weeks on new hardware and apporve it's use I'l start to have more faith in a global standard. Last I checked some of them were running yesteryears pc's..


For now I will stick to 30mins or 1 hour of OCCT, one thing I don't do however is force anybody to use what I use, nor do I diss them for not doing so.

Pedro stated his review was benchmark orientated, I stated that with the 3d benches I stuck up here @ XS too. The rest is 30 mins/1 hour OCCT stuff..


Back at my home site I had several very well respected friends make comparisons (they were all prime believers just like I was), with the Conroe architecture and a variety of motherboards, the conclusion was that OCCT was a harder hitting test that could be run for a third of the Prime slot for similar effect.

I experimented with this very heavily and found compatibility pertaining, a 30/60 mins OCCT run could be regarded as truly stable for a working PC. Again opinions will differ, so why should I force mine upon anyone else?

I would still be foolish if I thought everybody sees things that way, or more foolish still if I believed that this will be the case for all upcoming hardware and fully compatible everywhere, that has been my point. Do I run stress testing for a new board when posting in appropriate places? Yes I do, thus far I have had not one person come back to me and say well that 30mins or 1 hour OCCT run, I can't run my PC in it's daily tasks there it just won't do it. The most any one person has said is, 'why is this not 8 hours of prime?', Well why does it need to be if the pc runs everything you throw at it without fail? How often in Prime or any other tool for that matter checked for absolute consistency across CPU's, and memory combo's, a search can reveal incompatibilities at times...


Let people run what they will, myself personally, I won't have blind faith until I see more frequent checks or statements from the coders that they at least have checked their software for possible incompatibility or at least run the latest hardware to make sure... Or until I see that a 30 mins/1 hour OCCT test is truly inadequate for day to day tasks on a PC..


see ya
Raja