Overclocking CPU/GPU/Memory Stability Testing Guidelines

[daddylongs]

unsurprisingly, my chip seems to behave very similarly to yours Idontcare. I have been having a hell of a time trying to find a version of AI Suite II that will install on my machine.

(Asus p8p67 deluxe, win 7 x64)

Here is my current high stable point: taken while running BF3 beta:

 

[lehtv]

Of course with Sandy Bridge you don't have to disable power-saving features in order to overclock. This means that unless you are actively running CPU intensive tasks (i.e. benchmarking, rendering, transcoding, etc.), the temps are going to be only a little above ambient.

When I was using my Thermalright HR-02, I could shut off the fan for general computer use even though the 2600K was overclocked to 4.7 GHz.

So yes, you can cool it passively, but you might as well have a fan in there anyway. That way you can overclock higher and just turn the fan on when needed.

I think I'm going to be using my Mugen 2, along with low-RPM case fans... Mugen 2 isn't quite as good as HR-02 but it should do the job given how low-TDP Ivy Bridge will be. And I don't want to have to plug the fan in every time I actually use the CPU for something intensive. Too much hassle :/

 

[daddylongs]

Would you guys have any idea about why the CPU voltage reported in CPUz might be different than what I am trying to set it at in the BIOS or in the OS utilities (AI Suite II, Turbo evo) (Asus p8p67 deluxe)

 

[daddylongs]

Another day, another round of data to report: AVX does make a big difference!

Data point on AVX results, only difference between these two runs was AVX or not. Results:

LinX:
No AVX: 51.2 Gflops, Tmax = 62
Yes AVX: 100.6 Glops, Tmax = 67

Quite a difference there, thanks for the nudge.

 

[Idontcare]

Would you guys have any idea about why the CPU voltage reported in CPUz might be different than what I am trying to set it at in the BIOS or in the OS utilities (AI Suite II, Turbo evo) (Asus p8p67 deluxe)

CPUz Voltage = ROUNDDOWN((BIOS_Vcc/0.001)/8,0)*8/1000

Another day, another round of data to report: AVX does make a big difference!

Data point on AVX results, only difference between these two runs was AVX or not. Results:

LinX:
No AVX: 51.2 Gflops, Tmax = 62
Yes AVX: 100.6 Glops, Tmax = 67

Quite a difference there, thanks for the nudge.

:thumbsup: Many thanks for contributing this!

 

[lehtv]

CPUz Voltage = ROUNDDOWN((BIOS_Vcc/0.001)/8,0)*8/1000

Thx i've been wondering this too. Was kinda confused by that first but figured it out then ;P

My BIOS_Vcc = 1.19125V. CPU-Z says 1.184V on idle which is correct according to the above. But at load, it says 1.200V. Is this an actual voltage adjustment the CPU makes and should it happen...?

 

[Idontcare]

Thx i've been wondering this too. Was kinda confused by that first but figured it out then ;P

My BIOS_Vcc = 1.19125V. CPU-Z says 1.184V on idle which is correct according to the above. But at load, it says 1.200V. Is this an actual voltage adjustment the CPU makes and should it happen...?

I'm sure there are infinitely better ways to express it, that's just the one I've found to work for the hack/grunt that I am

Pedantic coders will descend to correct the errors in my noob coding ways in...3...2..1...

The 1.200V at load is correct to the formula as well, you can reverse it to then find out what your BIOS is attempting to set the Vcc to such that it drove the CPUz reading to a value of 1.200 as the VID and LLC changes kick in. (note, for 1.200V CPUz that also means 1.200V of Bios Vcc)

edit: just to clarify, when I say "BIOS Vcc" I do not mean the set value for Vcc in the BIOS, I mean the Vcc value that the BIOS is reading/monitoring...a value which naturally drifts as the load on the CPU changes. Your set Vcc in the BIOS can be an identical reading in those cases where the Vdrop and Vdroop are zero (thanks to LLC and/or measurement error)

 

[lehtv]

Thanks :thumbsup:

 

[NoobyDoo]

"OCCT Perestroïka v4.0.0 Available !"
http://www.ocbase.com/perestroika_en/index.php?News

Changelog (Since 3.0.1) :

MAIN

Brand new interface with real-time graphs !
Infinite mode is now used as default
In automatic mode, the IDLE periods are fully customizables
GPU:Memtest isn't available in this release (It is scheduled for complete revamp soon)
Settings simplified
Screenshot button (PNG format)

CPU:OCCT

No limit to the number of cores supported

CPU:LINPACK

Updated to support the newest Intel instructions

GPU:3D

Waiting also for complete revamp (already started - will be DirectX10 based (will run on a vanilla seven and vista). I'm studying DirectX11 (waiting for a card that supports it in fact)

MONITORING

Huge update to the monitoring and system information engine (supports Sandy Bridge and Bulldozer)
Fully customizable real-time graphs (look in OCCT's settings)
Minimum and maximum alert values can be set for every sensor !

INSTALLER

Will install .net framework 2.0 if not available
Updates DirectX9 if not up-to-date

TRANSLATIONS

New engine for translating OCCT - XML Based
Older translations had to be dropped - sorry
Instructions on how to translate OCCT will come soon...

There are too many features in v4 to be all written here in fact...

 

[Shmee]

very cool! I was also wondering, does it matter what res is used with GPU test? is default fine or should it always be run fullscreen native res?

....

well, i gave it a try, enabled error checking and fullscreen for GPU test, shader complexity 8, 1024x768 @ 60Hz and results:

seems like only one GPU was used. GPU1 maxed at about 68C. started out with high fps, then dropped to about 12 and stayed there. I suspect power throttling perhaps?

This is on unlocked 6950 xfire @ 840/1250

 

[Idontcare]

"OCCT Perestroïka v4.0.0 Available !"
http://www.ocbase.com/perestroika_en/index.php?News

Thanks NoobyDoo :thumbsup: OP link updated!

very cool! I was also wondering, does it matter what res is used with GPU test? is default fine or should it always be run fullscreen native res?

....

well, i gave it a try, enabled error checking and fullscreen for GPU test, shader complexity 8, 1024x768 @ 60Hz and results:

seems like only one GPU was used. GPU1 maxed at about 68C. started out with high fps, then dropped to about 12 and stayed there. I suspect power throttling perhaps?

This is on unlocked 6950 xfire @ 840/1250

Shmee, I forget the specific name of what OCCT "ramps" to increase the difficulty (shader depth, maybe ) but what you describe in terms of fps going from high to low is to be expected in my experience as the program ramps the rendering challenge for the card to get it hotter and hotter.

If you can only get OCCT to run on 1 GPU, can you run 2 instances of OCCT and have the second instance load the second GPU? (purely speculating on my part, I know nothing about these sli/xfire setups)

 

[AdamK47]

Wow! The GPU stress test is fixed. It REALLY does stress my GTX 580s now. Much more than what I've seen with folding. The fans were spinning at near 100% 5 minutes into the test.

 

[AdamK47]

If you can only get OCCT to run on 1 GPU, can you run 2 instances of OCCT and have the second instance load the second GPU? (purely speculating on my part, I know nothing about these sli/xfire setups)

Yeah, SLI/Crossfire doesn't work that way. See post #13.

 

[Shmee]

hmm, weird, checking with GPUz, max GPU load on GPU0 is 18%, while on GPU1 it is 0% :C

Am I power throttling?

....

Well, afterburner and kombustor make GPU0 hot, but not GPU1. Seems like the crossfire thing isnt working...

 

[AdamK47]

It could be power throttling. This new version of OCCT puts some serious strain on GPUs. This may sound like an obvious question, but do you have Crossfire enabled in the control panel? Try something else that you know has a Crossfire profile and see if MSI Afterburner shows correct GPU usage. It's been a while since I used AMD/Ati cards, but I believe there is an option to enable the crossfire logo in the CCC during 3D rendering to show if Crossfire is working on not.

 

[Shmee]

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

 

[OVerLoRDI]

Thanks for these links. I used LinX to discover my overclock was not 100% stable. Looks like my 2500k needs 1.425 volts for 4.5 Ghz. My mobo is not that high end, so it seems like my vcore drops a lot under load, so I have to be over aggressive in the BIOS settings.

Maybe I'll consider a more high end mobo when Ivy Bridge drops.

 

[Idontcare]

Thanks for these links. I used LinX to discover my overclock was not 100% stable. Looks like my 2500k needs 1.425 volts for 4.5 Ghz. My mobo is not that high end, so it seems like my vcore drops a lot under load, so I have to be over aggressive in the BIOS settings.

Maybe I'll consider a more high end mobo when Ivy Bridge drops.

That does seem like a lot of Vcc for that clockspeed, what are your temps?

 

[Schmide]

If you click full screen in the GPU 3D, crossfire/SLI will work.

 

[campbbri]

Thank you so much for the post. I used to just run blend Prime 95 and Mem86+ and call it a day, your test series is stronger, quicker, and easier.

I would second what some others say that the minimum requirement should be closer to 25 passes and the gold standard something more like 100 passes. I'm reconfiguring my overclock settings and I've had failures at 4 passes, 16 passes and 19 passes depending on the setup. Not many data points, but that suggests to me it's likely to see failures after 20 passes for a marginally stable setup. Also, each pass is so quick it's not much of a burden to go past 20.

I do have one question. I pass with both Prime95 (large FFT) and LinX with my current setup (i7 920 clocked to 3.6 Ghz). Unfortunately the memory test fails before 100% coverage even if I dial it down to stock clock, loosen the timings, and add memory voltage. Can I reasonably say the problem is the RAM modules and not the IMC or some motherboard component? I'd like to just replace the RAM sticks for ~$40, but I've been running Folding@home 24/7 for a few years with temps around 70 degrees, and I wonder if I've weakened some other component.

 

[Kenmitch]

Intel Burn Test also can be used to fine tune your voltages. Besides finding vcore required to be stable at an overclock you can also use it to optimize your voltages. I first noticed this with my i5-750 overclocking adventure.

Example: Hypothetical voltages and Gflops used but you'll get the picture! It's best to do multiple runs before you determine if the xtra voltage is needed as the Gflops tend to increase a little bit after the initial run. Suggest a couple of runs and average your Gflops during each step.

2500K at 4.5ghz 1.34v vcore (IBT stable) 96Gflops
2500K at 4.5ghz 1.36v vcore (IBT stable) 102Gflops
2500K at 4.5ghz 1.37v vcore (IBT stable) 104Gflops

Guess the point is sometimes it's not best for performance to run your vcore at the bare minimum required to stabalize your overclock.

Besides doing max memory I also like to play around with custom settings. Example 500 runs at 128mb's or so. I feel that possibly it may give the CPU a more realistic day to day workout as most of us are not slamming all the memory at once.

 

[Kenmitch]

LinX/IBT Some may think this is excessive but I consider 20 runs to be the minumum for stability and 50 runs the gold standard (or you can call it platinum). I've had it BSOD at 24 runs.

I guess in the end it's just a matter of how you got to the point your at in the stress testing. If your only looking to see how low you can drop the vcore and keep LinX/IBT stable then more runs would be best. Using something like IBT can also help you tweak your voltages for optimal performance. Things like vcore can greatly influence the Gflop output. I've seen as much as a 10% gain in Gflops vs being just stable with IBT in the past. At what point does a person calls it stable will fluctuate from person to person. Guess my point is if you do 50 runs it might have crashed at 53 or if you did 100 runs it might have crashed at 112. Guess the same would hold true for all forms of stress testing. Guess a person has to determine what they think is stable and take it from there. Sometimes a little bump in a voltage may be required at a later time.

 

[Idontcare]

Wasn't sure where to put this but I figured this thread would be a good place for a post on the topic of the impact of CPU operating temperature and the minimum Vcc needed to maintain stability, contrasted with the resultant power-consumption of the CPU.

These tests were conducted with a 2600K and an NH-D14 (both lapped) using NT-H1 thermal paste.

What the graph shows us is nothing more than what many of us already know, but it does confirm our expectations with data.

Depending on my fan configuration, and the ambient temps, I can run the 2600K at 4.5GHz with as little as 1.294V (measured w/multimeter) while being LinX stable for at least >= 5 passes (roughly 1.5hrs) provided the temps stay below 69°C.

The power consumption for the CPU while loaded w/LinX is 128W at these settings.

If the CPU max temps go above 69°C then I have to bump up the Vcc one notch to 1.298V and the power usage takes a minor jump to 131W.

As the operating temperature continues to rise the power consumption also continues to rise (for deeper explanation for why this occurs see this thread), and eventually we get to an operating temperature for which even the 1.298V is no longer enough to ensure LinX stability (around 93°C for this setup).

At that point the CPU require yet another voltage bump, to 1.304V, and now the power consumption is around 146-147W. Nearly 14% higher power usage than the same chip operating at the same clockspeed but with a lower voltage and lower temperature when the cooling is superior (30°C cooler, 63°C vs 93°C).

Keeping your CPU cooler will enable you to attain the same clockspeed with less voltage, a double benefit to lowering power consumption which itself feeds back into the equation to help keep those temps lower.

 

[Tsumi]

Very nice, thanks for the guide!

I just have this one thing to say about LinX. How long it runs highly depends on how much memory you have. And the length doesn't seem to scale linearly with the amount of RAM you have. If I recall correctly, with my i7 920 at 4.2 ghz and 12gb of RAM, it takes about 20 minutes to run one test. With 24gb of RAM, it takes about 50 minutes. So, to say that there is one set number for the amount of tests you should run doesn't make all that much sense. I believe it is more on how long LinX should be run, like P95.

 

[Idontcare]

Very nice, thanks for the guide!

I just have this one thing to say about LinX. How long it runs highly depends on how much memory you have. And the length doesn't seem to scale linearly with the amount of RAM you have. If I recall correctly, with my i7 920 at 4.2 ghz and 12gb of RAM, it takes about 20 minutes to run one test. With 24gb of RAM, it takes about 50 minutes. So, to say that there is one set number for the amount of tests you should run doesn't make all that much sense. I believe it is more on how long LinX should be run, like P95.

LinX/IBT only performs the error checking for stability purposes at the end of every cycle.

If you don't iterate through X-cycles then you have not stress tested your system to see if it is capable of repeatedly calculating the same set of computations X-number of times.

Precisely because the time involved per cycle is memory dependent (depends on how much ram you have installed and are testing), there is every reason to carry out the stress testing on a cycle-basis versus a time-basis.

The very same reasoning why you carry out ram stability testing on a cycle-basis (times through the loop test) which is also called "% coverage".