Overclocking CPU/GPU/Memory Stability Testing Guidelines

[geokilla]

Wow! I have the same problem. Recently upgraded my processor, could not activate LinX. I kept it running for more than 10 hours, never stressed my CPU.

Same problem here! Can't seem to stress test using LinX!

 

[crownjules]

Same problem here! Can't seem to stress test using LinX!

I had this problem and figured it out. ASRock has an OCing program in Windows and it was overriding the BIOS settings. So while the BIOS was set to a 33x CPU multiplier, the program was only set to 16x and therefore it never seemed like my CPU was getting worked.

So either uninstall/deactivate any Windows based OC program your motherboard runs or set it to the appropriate setting.

 

[geokilla]

I had this problem and figured it out. ASRock has an OCing program in Windows and it was overriding the BIOS settings. So while the BIOS was set to a 33x CPU multiplier, the program was only set to 16x and therefore it never seemed like my CPU was getting worked.

So either uninstall/deactivate any Windows based OC program your motherboard runs or set it to the appropriate setting.

I don't have any Windows based OC programs. It just shows 0% usage. I ended up using OCCT... Good enough.

 

[MPiland]

Also, don't forget after running these tests to check your event viewer for any critical errors that these tests aren't reporting. Don't want a silent corruption

 

[Monte45]

Awesome thread always known about OC'n capabilities but these programs make it simple for a new comer thanks for the info guys already love this place!

 

[moowarcow]

Hey guys,
I have a problem with Prime95 27.7 around the 1 hour mark I get error messages saying not responding. I don't get a BSOD. Also for IBT, OCCT, and Memtest they have passed the gold standard. Are there reasons why Prime95 keeps getting not responding?

 

[Idontcare]

Hi moowarcow, welcome to the forums

When you run prime95, are you running the "small FFT", "large FFT", or "blend" torture test?

Regardless your answer, the most likely reason for the program locking up is that your system isn't stable. Your answer to the above question will provide insight into which component in your system is the likely culprit.

The fact that the program locks up but your rig doesn't BSOD just means your rig is likely to be almost very nearly stable, it is just ever so slightly unstable and probably just needs a minor voltage bump to establish robust stability.

Have you tried raising your CPU voltage to see if it then passes the prime95 test?

 

[moowarcow]

Hi moowarcow, welcome to the forums

When you run prime95, are you running the "small FFT", "large FFT", or "blend" torture test?

Regardless your answer, the most likely reason for the program locking up is that your system isn't stable. Your answer to the above question will provide insight into which component in your system is the likely culprit.

The fact that the program locks up but your rig doesn't BSOD just means your rig is likely to be almost very nearly stable, it is just ever so slightly unstable and probably just needs a minor voltage bump to establish robust stability.

Have you tried raising your CPU voltage to see if it then passes the prime95 test?

Thanks for the quick reply!

I followed the guide using the Large FFT settings. I've used blend before it passed 12hrs+ without any issues. I will try to bump the voltages up a little and see if that makes any difference.

 

[Idontcare]

Large FFT giving errors is usually indicative of a memory controller instability (too low on the VDimm or VCSSA). So you might try bumping those up a notch or two just to see if it remedies the issue with prime95 locking up.

And just to confirm, you have tried lowering the actual overclock itself (reduced the multiplier) and you know Prime95 runs stably on your rig at the lowered overclock?

 

[muk1604]

thx

 

[BonzaiDuck]

I'll post this here, to avoid the clutter of a new thread.
I've been fine-tuning my i7-2600K in a P8Z68-V Pro motherboard. Cooling: Noctua NH-D14. Over-clock objective: 4.6 Ghz. After a year with this machine, I felt that my VCORE was just a tad excessive.

I followed a plan to keep the Offset at either -0.005V or +0.005V, adjusting the "V_Turbo" or extra voltage in turbo-mode exclusively. Changing between the negative and positive offset values allowed me to make adjustments of only 2 mV -- as opposed to the 4mV adjustments in V_Turbo.

After reaching a point where a core fails (without BSOD) in Prime95 sFFT or lFFT after 8.5 hours, I decided to kick up the LLC setting to "High" or "50%" and then compensate with the voltage setting to get me just short of a 5mV increase over the previous setting. Then I ran LinX, choosing "Affinity" to four threads with HT Enabled.

This is what I got for the initial V_Xtra of 4 mV:

Now, I kick up the VCORE ("Extra Turbo voltage") another 0.004V to 0.008V:

Suddenly, the grouping is "tighter," but I know that an i7-2600K @ 4.6 Ghz should show as much as 127 GFLOPS, so I kicked up the voltage again by another 0.004V:

Later, I tried to replicate this setting for the same results, but they looked more like the previous one. I kicked up the voltage another 0.004V notch, but the results looked again like the previous one.

I realize that background processes will reduce the GFLOPS result; there are many things going on. Also, I'm not sure what I should expect: should there be a distribution of GFLOPS values? Should they be consistently near the same?

All comments will be appreciated.

Thanks.

 

[Idontcare]

BonzaiDuck, the errant behavior you are witnessing is because of core parking, your OS is changing which cores are the parked cores but the affinity assignment prevents the LINX threads from migrating to the unparked cores.

It has been my experience that Windows is random in when/how/why it parks cores when HT is enabled. You have to keep an eye on it. I guarantee you on those runs which turned in 80GFlops your CPU utilization had dropped from 50% to 38% because one of the cores changed its parked status during the run.

As for the variation - you appear to not be locking/specifying the amount of memory used from test to test. Make a note of the specific problem size you tend to use and then make a point of always specifying that specific problem size in all your future tests. The GFlops value is problem problem size specific.

For 16GB with Win7 Ult x64, I use problem size 43122.

 

[BonzaiDuck]

Heh -- getting older and losing my "edge." I missed a step in the in the LinX setup, restricting the threads to 4.

Now I'll have to rerun everything. Maybe I could still get robust data from small samples of 10 iterations.

But was I in the ballpark here? Is some of what I'm seeing so far suggesting an approach to fine-tuning VCORE/related settings?

 

[Idontcare]

But was I in the ballpark here? Is some of what I'm seeing so far suggesting an approach to fine-tuning VCORE/related settings?

You do have to check your GFlops even when stable to ensure you aren't on the hairy edge of only being Linx stable by virtue of ECC saving your computational bacon.

The way you determine if this is the case is once you are seemingly "linx stable" you then bump up the voltage by a non-trivial amount (say 0.05V) and rerun the tests. Of course your temps will go way up, but that isn't the point.

The point is your GFlops should remain unchanged. If they improve (higher GFlops) then you really weren't stable at the lower voltage, ECC was saving your bacon.

 

[BonzaiDuck]

You do have to check your GFlops even when stable to ensure you aren't on the hairy edge of only being Linx stable by virtue of ECC saving your computational bacon.

The way you determine if this is the case is once you are seemingly "linx stable" you then bump up the voltage by a non-trivial amount (say 0.05V) and rerun the tests. Of course your temps will go way up, but that isn't the point.

The point is your GFlops should remain unchanged. If they improve (higher GFlops) then you really weren't stable at the lower voltage, ECC was saving your bacon.

It seemed your data on the i7-2600K would've got you past that with your 127 GFLOPS result? I somehow think that an incremental process will work for me here . . . I suppose if there's too much change in mean values, I could try bigger increments . . . . .

The CPUs will differ. You may have tested for only 5 passes Linx stable.

I'll just have to get through it, one way or the other, I suppose . . . But I'm thinking I could reduce this to a 5-pass sequence, because these numbers look so tight . . .

Seems I'd already been up and down the voltage ladder with this chip, testing in other ways . . . We'll see . . .


UPDATE: I've now dropped the voltage by 8 mV, and still the majority of readings in samples of 10 are >= 129. Or -- the average hasn't changed going up 8mV, and going down by that amount means the average decreases by maybe 1 GFLOP.

 

[BonzaiDuck]

I'll go ahead and post this, with the assertions it contains. As IDontCare observers, the GFLOPS results in LinX stability testing give one an idea about whether the CPU is losing clock-cycles through error-correction in the cache. The more clock-cycles lost, the lower the GFLOPS or the departure from the maximum you would expect. Insufficient voltage would create such a condition.

Also as IDontCare notes, you need to find the maximum GFLOPS your CPU can reach for the particular clock speed.

Once done, you can go through your voltage-trimming exercise and stability testing, and use several LinX passes to find a point where most sample observations -- the GFLOPS for a string of iterations or runs -- appear close to the maximum.

As you decrease voltage from that point, a declining number of GFLOPS will be close to the maximum. The mean GFLOPS for a given sample will fall below the maximum. And the standard deviation, "range," or variation around the mean will increase.

As you approach the point of sufficient voltage, more and more sample observations will be close to the maximum; the mean GFLOPS of the sample will be closer to the maximum; and the variation will attenuate to the point of being insignificant.

That's what I have observed, anyway . . .

 

[Momar99]

not impressed by TurboBoost 2.0? haha

 

[nxttruong]

What a huge store of information! This is the first time I have known these sharings.

 

[MsLuck]

update, seems to stress both cards when I OC, but not at stock clocks. Yes, crossfire is enabled. How odd, though.

What do you mean by "odd"?

 

[blastingcap]

IDC do you have recommendations for stresstests running under Linux Ubuntu and how many times/how long to run them for, by any chance?

This thread is a work in progress. The OP will be updated/amended to reflect information as made available by thread contributors.



LinX (Intel burntest) is superior to Prime95 small FFT for determining CPU core logic stability.

• Must run with the IBT thread count set equal to the physical core count of the CPU.
• HT slows it down and reduces the ability to determine stability. Set to 4 threads on a 2600K.
• Set memory to "All".
• Stability Criterion: Must pass 5 cycles minimum, passing 20 cycles is preferred (considered gold standard)

Prime95 large FFT is superior to LinX for determining L3$/IMC stability.

• Must use large FFT, blend is insufficient. <- reports indicate this is false for AMD stability tests, see post #4
• HT is ok for this test.
• Stability Criterion: Must pass 2 hours minimum, passing 12 hours is preferred (considered gold standard)

HCI memtest is superior to LinX, Prime (large or small), and memtest86+ for determining memory stability.

• Launch one instance per thread supported by CPU (8 instances of HCI memtest for 2600K)
• Set each HCI instance to use an appropriate fraction of the memory...16GB on a 2600K means each HCI instance (there will be 8 instances) are to use 2048 MB.
• Stability Criterion: Must pass 200% coverage minimum, passing 1000% coverage is preferred (considered gold standard)

OCCT GPU test w/error checking enabled is superior to Kombustor for determining GPU stability. (updated link to OCCT 4.0.0, thanks NoobyDoo!)

• Error checking MUST be enabled by the user (check the box), otherwise you are leaving it up to your eyes to detect visual artifacts which renders the test entirely subjective.
• Stability Criterion: Must pass 20 minutes minimum, passing 1 hour is preferred (considered gold standard)

 

[Idontcare]

IDC do you have recommendations for stresstests running under Linux Ubuntu and how many times/how long to run them for, by any chance?

Unfortunately I do not

But if you do, or find yourself pulling such a list of recommendations together, I'd be more than happy to include it in this thread's OP for everyone's benefit.

 

[Ed1]

I just tried this with 8gig memory and 3570k (3.8ghz) I got 98.7 avg (mid to high 98's)Gflops .

Is there a DB of what is normal range, does it depend on clock speed mem speed or all above ?

Also I noticed there some newer versions (LinX(linx(0.64)_10.3.7.12.2) on the web is there reason the one in first post is best ?

Edit: I see value goes up as you raise memory size, I think I had it on 1024 or so, I raised to 4096 and it went up to 105 Gflop .

 

[Idontcare]

I just tried this with 8gig memory and 3570k (3.8ghz) I got 98.7 avg (mid to high 98's)Gflops .

Is there a DB of what is normal range, does it depend on clock speed mem speed or all above ?

Here are my results. Yes GFlops goes up with memory, as does power consumption and heat (temperature).

Basically I get 28.412 GFlops/GHz with my 3770k.

Also I noticed there some newer versions (LinX(linx(0.64)_10.3.7.12.2) on the web is there reason the one in first post is best ?

There is nothing special about the one linked in the OP, at the time it was the one version I found that had been updated to include AVX so that is why I linked it.

If you know of a more updated/recent version and have the download link for it then please post it here in the thread and I'll update the OP with it as well.

 

[Ed1]

Here are my results. Yes GFlops goes up with memory, as does power consumption and heat (temperature).

There is nothing special about the one linked in the OP, at the time it was the one version I found that had been updated to include AVX so that is why I linked it.

If you know of a more updated/recent version and have the download link for it then please post it here in the thread and I'll update the OP with it as well.

thanks so I am about in ballpark I think (will have to try with all mem size) .

The thing is I didn't try links , as I don't know contents but here google search limk with many of them .

https://www.google.com/#hl=en&sugex...ih=829&bav=on.2,or.r_gc.r_pw.r_cp.r_qf.&cad=b

here LinX 11.0.0.004 version listed (I DL and seemed to work ok )
http://forums.tweaktown.com/gigabyt...system-info-benchmarking-stability-tools.html

Hold on, something funny is happening with this version . not sure if its because I have older version in folder up one level but this version when you hit start it opens a new version running and the orginal version doesn't do anything, even though it appears to be running . weird

 

[maleko]

i may have missed it, but does anyone have any recommendations on how i might do stress testing without a windows install or even a hard drive? a live cd version of prime95 would be perfect, but i cant seem to find one.

i'm working on doing a build/rebuild of mine and my wife's computers and i'd like to test them before i toss win7 back on them