Ivy Bridge stability testing

[coffeejunkee]

I've been overclocking for quite some time now and always had good results using the common stresstesters for stability testing. During the Core2 years I mainly used Prime95 and for Lynnfield I switched to LinX which worked even better and quicker. But my current IVB cpu is a different animal. I have used Prime95 (with a variety of settings), LinX, IBT, OCCT, Aida64, BurninTest and Intel XTU but none of them were anywhere near succesful in determining stability. At settings that would run Prime95 overnight and 100 runs of LinX I still had crashes in some games, not to mention the many WHEA errors.

So I mostly relied on playing games to test for stability but having to play for hours to catch some instability isn't very practical. So I searched some more and found this post: http://www.overclock.net/t/1338762/...o-prime95-for-stable-testing/10#post_20119984. I decided to try it out and I had some good results using this method. I downloaded Arena 3.0 from http://www.playwitharena.com/?Download, Houdini 1.5 from http://www.houdinichess.com/ and Critter 1.6 from http://www.chess.com/download/view/critter-16. After downloading you'll first have to install the engines under engines > install new engine and then load them as white and black players under engines > load engine. Then click demo to have them play a match against each other.

So, can your overclocked Ivy (or Sandy) cpu have these engines battle it out without bsod's, error messages or whea errors? How about when you're doing some heavy duty browsing in the meantime? Or maybe you have yet a different way of testing for stability?

 

[Idontcare]

There's 2000 instructions in your IB processor. Some will function fine at 6GHz, others won't produce correct mathematical results if clocked at say 4GHz unless you boost the voltage above stock.

The point is that no two IB processors are the same. The weakest link in stability will be different between just about every enthusiast out there that is OC'ing.

Since no commercially available freeware program will test the mathematical correctness of every single instruction in the ISA, we have no way of knowing which particular instruction is the weakest link in our efforts to OC.

For multi-core CPUs this becomes even more complicated because the weakness will vary from core to core as well as instruction to instruction.

If your specific IB processor happens to have a particularly weak circuit that corresponds to a specific instruction used in say Prime95 then Prime95 is going to magically seem like the best stress tester you could ever use (and for your specific processor it may well be true)...but the guy next to you may find the weakest circuit in their specific IB is in an instruction that is only stressed heavily while playing BF3 and so BF3 is going to seem like the magic bullet in optimizing the stability and OC for their specific IB CPU (and it will be, for their specific CPU).

Whea errors, BSODs, etc are all sides of the same multi-faceted "weakest link" issue. Each is the result of a different combination of instructions being stressed and they either compute a mathematically correct output or they compute junk results and lead to instability. No two processors are going to be the same in this regard, but any given processor is susceptible to the problem.

That is why people call it the silicon "lottery". There is an element of randomness in the equation owing to entropy and the manufacturing process (process variation).

This is why single-faceted OC guides fall flat. There is a distribution of processors out there, over-laid by a distribution of software packages, and the mesh between the two produces a spectrum of corner-cases where the shmoo plot goes from green to red at any given clockspeed, voltage, and operating temperature.

 

[Concillian]

Yeah. you'll find there are varying definitions of stable on the internet.

Me, I do IBT AVX version 20 runs without getting closer than 25 C distance to Tj. Then at that clock and voltage, Prime 95 blend and Small FFT 12 hours each. Then I turn the clocks down 5% without touching the voltage and call that stable.

There's too much the stability testing programs will miss and I value stability. I've not had a CPU have ANY issues with that method, but it doesn't give any bragging rights either.

I learned long ago (Celeron 300ish timeframe?) That stability tests are just a guideline and don't necessarily represent actual stability. Margin for error is a good thing... unless you're just OCing to post screenshots on the internet... then it's really bad.

It's like when I do fuel / timing / manifold pressure tuning on a car. You can't go in with a horsepower goal for a given setup. You have to listen to what the car is saying to you and take it where it leads you. That will get you the best HP you can get for any given level of margin. You can't force what's not there.

 

[Zanovar]

Great stuff idc as per

handbrake and similar progs are giving me the whea errors.stable in most things i guess(this is ivy btw)

Btw coffee tried the chess match,was quite the match

stick to 4.6 safe i guess ha

 

[coffeejunkee]

Idontcare: So we'll never know true stability (Unless we work at Intel or AMD)? There could always be some program out there that finds our weakest link...Anyway, I understand the silicon lottery and I might have a slightly finicky part but still, it seems to me these modern cpu's are getting really complex and just running Prime95 isn't enough anymore, whereas earlier that might weed out like at least 95%. I could be wrong, but that's what I would say based on my own experience + reading about a couple other people having a similar one (obviously the guy in the post I linked, and also some on these forums). In any case, I've become quite sceptical of people posting Prime95/Linx screenshots for 'proving' their overclock is stable.

Concillian: That sounds pretty conservative, but I can imagine that will be quite stable. Ivy likes to get hot though, I'd have to go from 4.6 to 4.2 to meet your limits.

Zanovar: Yeah, they're pretty good. I wouldn't want to play either of them. So what vcore you use for 4.6 (minimum)?

 

[Idontcare]

Idontcare: So we'll never know true stability (Unless we work at Intel or AMD)?

Correct.

There could always be some program out there that finds our weakest link...

That is the eternal hope we enthusiasts have, that something out there (preferably free at that) will be "good enough".

Anyway, I understand the silicon lottery and I might have a slightly finicky part but still, it seems to me these modern cpu's are getting really complex and just running Prime95 isn't enough anymore, whereas earlier that might weed out like at least 95%.

I agree there too. What has made things more complicated is the fact our CPUs are becoming more like SoCs with more and more stuff integrated onto the die, as well as more and more instructions being added to the ISA.

 

[Zanovar]

Idontcare: So we'll never know true stability (Unless we work at Intel or AMD)? There could always be some program out there that finds our weakest link...Anyway, I understand the silicon lottery and I might have a slightly finicky part but still, it seems to me these modern cpu's are getting really complex and just running Prime95 isn't enough anymore, whereas earlier that might weed out like at least 95%. I could be wrong, but that's what I would say based on my own experience + reading about a couple other people having a similar one (obviously the guy in the post I linked, and also some on these forums). In any case, I've become quite sceptical of people posting Prime95/Linx screenshots for 'proving' their overclock is stable.

Concillian: That sounds pretty conservative, but I can imagine that will be quite stable. Ivy likes to get hot though, I'd have to go from 4.6 to 4.2 to meet your limits.

Zanovar: Yeah, they're pretty good. I wouldn't want to play either of them. So what vcore you use for 4.6 (minimum)?

Well i thought i was stable at 1.286,that was before i started using handbrake and winx hd:S.seems stable at 1.3.the v is rising.

 

[2is]

There's no quick way to test stability. My own personal methods is to run IBT at high stress levels for 10 passes, this is just to test my cooling more than stability.

I then run through all my benchmarks, and run Prime95 Small and Large FFT for 24 hours.

But even then, sometimes that's not enough. I've passed all these tests and have still had something as seemingly simple as Internet Explorer crash from time to time until I raised voltage further or lowered frequency.

So really, for me it's a combination of running the known utilities we use as well as long-term usage of the machine itself.

Personally, I've opted for 4.2GHz on my 3770k. It's fairly mild but the CPU hit this speed quite easily with a modest increase in voltage and temperature. Going higher needs more voltage and/or enabling more exotic features like LLC or PLL Overvoltage and that's just not worth the extra couple hundred MHz I may be able to squeeze out of it. Especially if in the back of my head, every time I encounter an issue, I'm wondering if it's due to an unstable OC.

 

[coffeejunkee]

Well i thought i was stable at 1.286,that was before i started using handbrake and winx hd:S.seems stable at 1.3.the v is rising.

Seems to be about the same as my cpu, any less than 1.328 in cpu-z and I get whea errors. LinX runs happily at 1.280 though.

There's no quick way to test stability. My own personal methods is to run IBT at high stress levels for 10 passes, this is just to test my cooling more than stability.

I then run through all my benchmarks, and run Prime95 Small and Large FFT for 24 hours.

But even then, sometimes that's not enough. I've passed all these tests and have still had something as seemingly simple as Internet Explorer crash from time to time until I raised voltage further or lowered frequency.

So really, for me it's a combination of running the known utilities we use as well as long-term usage of the machine itself.

Personally, I've opted for 4.2GHz on my 3770k. It's fairly mild but the CPU hit this speed quite easily with a modest increase in voltage and temperature. Going higher needs more voltage and/or enabling more exotic features like LLC or PLL Overvoltage and that's just not worth the extra couple hundred MHz I may be able to squeeze out of it. Especially if in the back of my head, every time I encounter an issue, I'm wondering if it's due to an unstable OC.

So if it's sometimes not enough, why bother at all with the 24h Prime95 runs?

 

[2is]

24 hour prime runs alone isn't enough IMO, regular usage alone isn't either. It's the combination of all the tools together that works. I can just as easily seem stable and fail prime as I can pass prime and fail something else. There's no one magic bullet so we need to se a cocktail of tests.

Just like when a fighter gets ready for an upcoming fight, he needs a combination of diet, exercise, and training. Just because diet alone isn't enough doesn't mean you don't do it as its still necessary for success. Same goes for the training and exercise.

 

[Yuriman]

In my particular case I can run IBT or Prime for 48 hours @ 4.4ghz with no +v, but the emulator "Dolphin" crashes within 45 seconds. It takes +31mv before Dolphin will run over night.

Apparently whatever instructions it uses are a weaker link in my chip than those used in IBT. YMMV.

 

[24HZ]

You may not even be '100%' stable at stock speeds. Think about that. Stability needs a frame of reference. Is it stable for what you intend to do on it? Is it stable for every single pathway in the probability space? Is it stable with a bunch of synthetic tools that cover a large region of that space but never all of it?

Then there's also know and unknown bugs in ALL cpus made by anyone that will cause the 'bad things' to happen regardless of the speed. You just have to pray that whatever mission critical stuff you're running(if at all) doesn't do that.

 

[Zanovar]

24 hour prime runs alone isn't enough IMO, regular usage alone isn't either. It's the combination of all the tools together that works. I can just as easily seem stable and fail prime as I can pass prime and fail something else. There's no one magic bullet so we need to se a cocktail of tests.

Just like when a fighter gets ready for an upcoming fight, he needs a combination of diet, exercise, and training. Just because diet alone isn't enough doesn't mean you don't do it as its still necessary for success. Same goes for the training and exercise.

haha

 

[UaVaj]

only use prime95 (and those other similar programs) as a quick test to determine overclocking thresholds.

the true test of stability is actually playing a cpu intensive game (sc2, wow, bf3, crysis3) and not crashing.

----

i.e. prime95 stable all night however game would not last 1 hour before crashing.

 

[coffeejunkee]

24 hour prime runs alone isn't enough IMO, regular usage alone isn't either. It's the combination of all the tools together that works. I can just as easily seem stable and fail prime as I can pass prime and fail something else. There's no one magic bullet so we need to se a cocktail of tests.

Just like when a fighter gets ready for an upcoming fight, he needs a combination of diet, exercise, and training. Just because diet alone isn't enough doesn't mean you don't do it as its still necessary for success. Same goes for the training and exercise.

Diet, exercise and training are all proven to be effective. Prime95 (and other stresstests) I'm starting to seriously doubt if it is, seems like just a waste of time and energy to me now. Case in point my cpu, needing 0.04V extra for 'real' stability is a lot.

In my particular case I can run IBT or Prime for 48 hours @ 4.4ghz with no +v, but the emulator "Dolphin" crashes within 45 seconds. It takes +31mv before Dolphin will run over night.

Apparently whatever instructions it uses are a weaker link in my chip than those used in IBT. YMMV.

Any tips on using this Dolphin emulator?

You may not even be '100%' stable at stock speeds. Think about that. Stability needs a frame of reference. Is it stable for what you intend to do on it? Is it stable for every single pathway in the probability space? Is it stable with a bunch of synthetic tools that cover a large region of that space but never all of it?

Then there's also know and unknown bugs in ALL cpus made by anyone that will cause the 'bad things' to happen regardless of the speed. You just have to pray that whatever mission critical stuff you're running(if at all) doesn't do that.

I've thought about that, but I can't do much more than trust Intel. On that note, some say 'you must' Prime for 8h, some say 12h, others even 24h...where does it stop?

only use prime95 (and those other similar programs) as a quick test to determine overclocking thresholds.

the true test of stability is actually playing a cpu intensive game (sc2, wow, bf3, crysis3) and not crashing.

----

i.e. prime95 stable all night however game would not last 1 hour before crashing.

I consider gaming to be the best test as well, but it takes a lot of time, especially if you want to do various overclocks like me. I used to play Forged Alliance with 7x cheating ai but sometimes it took hours to crash. Something a little less time intensive would be nice, and I think the chess match does a good job.

 

[2is]

Diet, exercise and training are all proven to be effective. Prime95 (and other stresstests) I'm starting to seriously doubt if it is, seems like just a waste of time and energy to me now. Case in point my cpu, needing 0.04V extra for 'real' stability is a lot.

As I said, I've had instances where I've appeared to be stable in everything but prime also. Based on my own experiences, its an effective tool just not on its own. That's no different from any other tool you decide to use, including gaming. It's proven to be an effective ingredient on all my builds. I start with a few hours of prime, and end with a 24 hr run after I've done all my other tests.

 

[coffeejunkee]

Well, I guess we'll have to agree to disagree. To me the point always was Prime95 puts such a heavy load on your cpu that it will easily and relatively quickly determine stability, thereby making other tests superfluous. But if it isn't effective at doing that I don't see the point since it's rather unrealistic compared to many real-life apps. Mind you, if this was 2007 I would say it's the best for stresstesting, but with 2013 cpu's I'm not convinced it is at all.

Anyway, have you tried the Arena test? Takes just 15-20 mins.

 

[Concillian]

It wasn't enough in 2007 either. In the Barton core AthlonXP days, I had an OC that would pass 24 hours of Prime95, Small FFT / Large FFT... didn't matter, It would pass it. But if I launched a certain application I would reliably blue screen in less than 1 minute. I forget what that application was... I think it was a game.

I don't remember if I stabilized it with voltage or lower clocks. Doesn't really matter, the point is that Prime95 was never enough on it's own. That experience is when I started backing off clock by 5% at same voltage from wherever my stress tests were stable.

 

[2is]

Well, I guess we'll have to agree to disagree. To me the point always was Prime95 puts such a heavy load on your cpu that it will easily and relatively quickly determine stability, thereby making other tests superfluous. But if it isn't effective at doing that I don't see the point since it's rather unrealistic compared to many real-life apps. Mind you, if this was 2007 I would say it's the best for stresstesting, but with 2013 cpu's I'm not convinced it is at all.

Anyway, have you tried the Arena test? Takes just 15-20 mins.

I have not. I'll give it a shot on my next overclocking venture though.

It wasn't enough in 2007 either. In the Barton core AthlonXP days, I had an OC that would pass 24 hours of Prime95, Small FFT / Large FFT... didn't matter, It would pass it. But if I launched a certain application I would reliably blue screen in less than 1 minute. I forget what that application was... I think it was a game.

I don't remember if I stabilized it with voltage or lower clocks. Doesn't really matter, the point is that Prime95 was never enough on it's own. That experience is when I started backing off clock by 5% at same voltage from wherever my stress tests were stable.

Agreed, on it's own, it was never enough, nor was any one application. It was also never "quick" even when I was stress testing single core Socket A Athlons, a 24hr run was the recommended norm.

 

[Concillian]

Agreed, on it's own, it was never enough, nor was any one application. It was also never "quick" even when I was stress testing single core Socket A Athlons, a 24hr run was the recommended norm.

Yeah, SuperPi was quick and would generally show you RAM / FSB / BCLCK instability very quickly. You had to give that margin too, but it was great quick and dirty for finding a rough estimate of the limit. Prime95 has never been quick. Often I've seen issues that take 4+ hours to show up.

Now IBT does both. Clock the CPU down and it's good at finding RAM / FSB / BCLCK instability. Give the RAM / FSB margin and clock the CPU back up and find it's limit at your comfortable voltage. Back down clock for margin and do the same for the uncore. Bam done. What used to take me a week or two with Prime 95 now takes 2 evenings.

 

[PhIlLy ChEeSe]

only use prime95 (and those other similar programs) as a quick test to determine overclocking thresholds.

the true test of stability is actually playing a cpu intensive game (sc2, wow, bf3, crysis3) and not crashing.

----

i.e. prime95 stable all night however game would not last 1 hour before crashing.

And was there any noticed game improvements because of the over clock, IE frame rate etc? The OP seems to be asking that question indirectly, someone else posted the answer anything over 4.6 the gains fall off quickly.

 

[CTho9305]

Idontcare: So we'll never know true stability (Unless we work at Intel or AMD)?

Even the manufacturer doesn't always know. Sometimes they'll get problem reports from a customer or parts get returned, and they'll end up discovering a slow path in the chip that they weren't expecting. At that point they have to update their test programs to cover that new path. Generally this is more of a problem early in the lifecycle of a product.

 

[UaVaj]

And was there any noticed game improvements because of the over clock, IE frame rate etc? The OP seems to be asking that question indirectly, someone else posted the answer anything over 4.6 the gains fall off quickly.

games that were cpu heavy showed considerable improvement.

games that were gpu heavy showed slight improvement.

 

[Concillian]

For games, benefits fall off a lot earlier than 4.6 GHz, IMO. Like this thread where the guy doesn't notice a difference playing BF3 on a Gulftown vs a Haswell even though one shows a significantly higher FPS.

http://forums.anandtech.com/showthread.php?t=2325110

Real world usage gave a 33% increase from Gulftown to Haswell. However, both of these systems are so fast that it practically doesn't make any difference. In regular usage, you couldn't tell them apart. In gaming, the difference is too slight to notice. They are both more than adequate for a high level of play.

At some point there's simply "enough" CPU. In that case he doesn't mention Overclocking, so the Gulftown i7-970 is essentially acting like a Haswell clocked in at well under 3 GHz.

"Enough" happens at different MHz on different games, but for all but a very few select games, even a stock Haswell k series is "enough" to have "perfect" play where you won't notice another 30% FPS increase.

 

[2is]

"The guy" is a key phrase there, meaning exactly 1 person. There are a lot of other individuals who would notice. I'd say most PC gamers could easily appreciate the 33% improvement. Yes, there is a point where more performance doesn't really do much. 33% from Gulftown to Haswell in BF3 is far from being it IMO.