Overclocking CPU/GPU/Memory Stability Testing Guidelines

[Idontcare]

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&#37; 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)

 

[Avalon]

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.
• 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.

• 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)

I kind of feel that should be a sticky. Was unaware of HCI. Thanks!

 

[ThatsABigOne]

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.
• 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.

• 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)

:thumbsup: This should be a sticky! I did not know about those 2 stress testers!

 

[sangyup81]

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.
• 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.

• 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)

Like others have said, sticky please!

About the Prime95 Large FFT, I've gotten crashes on my Phenom II on Blend that I haven't gotten on Large FFT. This required me to run it 24 hours with crashes happening on the 20th hour a lot of times. These crashes were fixed by increasing increasing CPU or CPU-NB volts or downing the CPU or CPU-NB clock speeds.

 

[Idontcare]

Like others have said, sticky please!

About the Prime95 Large FFT, I've gotten crashes on my Phenom II on Blend that I haven't gotten on Large FFT. This required me to run it 24 hours with crashes happening on the 20th hour a lot of times. These crashes were fixed by increasing increasing CPU or CPU-NB volts or downing the CPU or CPU-NB clock speeds.

Noted in the OP.

Typically, issues that "blend" would capture will also be more readily detected with HCI memtest as the difference between blend and large FFT is the memory usage itself.

It would be nice if we could get confirmation of blend being superior to large FFT on Intel platforms as well...then we could just wholesale adopt the "use blend" proposal an drop the "use large FFT" recommendation entirely.

Anyone have anecdotal evidence to support such a proposal?

 

[daddylongs]

Thanks for putting all this amazingly helpful info in one place!

Do you have thoughts on Real Temp vs. Core Temp 1.0?

My next steps are to see how low I can get my CPU Voltage at 45x100. Currently I am stable at 1.30v. Do you have suggestions on what increments I should be lowering at? .05v?

Asus P8P67 Deluxe -- SB 2600k -- Mugen 3 (SCMG-3000) -- DDR3-2000 (Gskill Trident 2x2GB)
Antec 300, with 2x extra Scythe 120mm 900rpm front fans. -- OCZ Vertex3 240GB -- evga GTX 470 superclocked

 

[thegimp03]

Thanks for the information. I was unaware of those programs outside of Prime95 and memtest86+, but don't really consider myself in-the-know on these things anyways as I only build a system maybe once every 3-5 years.

 

[Idontcare]

Do you have thoughts on Real Temp vs. Core Temp 1.0?

I use both.

Real Temp has some nice feature but it only works with the Intel CPU's as I understand it.

Core Temp has nice temperature logging features. Real temp will also log temps but it appends each session's worth of data into one big mega-file whereas Core Temp will save log files by timestamp of the session's start which is nice for the purpose of going back to lookup the results of bench sessions.

My next steps are to see how low I can get my CPU Voltage at 45x100. Currently I am stable at 1.30v. Do you have suggestions on what increments I should be lowering at? .05v?

My mobo lets me change Vcc in 0.005 volt increments, which is what I use.

But I iterate with more of a "rapid find" approach by dropping launching a session of LinX, let the CPU get good and warm at my target GHz, and then using AI Suite I start dropping the Vcc in 0.005 increments. Let it stable, wait maybe 30 seconds, then drop again, wait 30s, drop again. All while LinX is running in the background.

Eventually LinX will either halt from a computation error or the system will BSOD, then I move up a couple notches and start a longer LinX run (5 cycles) without monkeying around with Vcc while it is running. If it errors then I bump up again.

For the most part, with my 2600K, in the 4GHz range for every 0.1GHz increase (or decrease) in speed I have to move the voltage by 0.02 Volts (more or less).

 

[Shmee]

any good stress tester for testing multiple GPUS at once with SLI/xfire?

 

[Idontcare]

any good stress tester for testing multiple GPUS at once with SLI/xfire?

Good question, I'm actually quite ignorant of OC'ing in the >1 video-card space but I am glad you brought it up because I aim to SLI my GTX460's and I was assuming OCCT was SLI-aware.

I take it that OCCT is not SLI-aware?

 

[daddylongs]

I use both.

Real Temp has some nice feature but it only works with the Intel CPU's as I understand it.

Core Temp has nice temperature logging features. Real temp will also log temps but it appends each session's worth of data into one big mega-file whereas Core Temp will save log files by timestamp of the session's start which is nice for the purpose of going back to lookup the results of bench sessions.



My mobo lets me change Vcc in 0.005 volt increments, which is what I use.

But I iterate with more of a "rapid find" approach by dropping launching a session of LinX, let the CPU get good and warm at my target GHz, and then using AI Suite I start dropping the Vcc in 0.005 increments. Let it stable, wait maybe 30 seconds, then drop again, wait 30s, drop again. All while LinX is running in the background.

Eventually LinX will either halt from a computation error or the system will BSOD, then I move up a couple notches and start a longer LinX run (5 cycles) without monkeying around with Vcc while it is running. If it errors then I bump up again.

For the most part, with my 2600K, in the 4GHz range for every 0.1GHz increase (or decrease) in speed I have to move the voltage by 0.02 Volts (more or less).

Awesome! I will do this after I install the new cooler. I'll make a public google doc and we can all add our data, and we can see a line for each persons CPU. I am curious about how mine lines up against others here, and also how wide the standard deviation is.

Do you know, in your rapid method, for each time you lower the Vcc, do you know if the temperature affects AI Suite's CPU speed that it settles on? I am trying to figure out what differences in CPU cooling/temperature would have on that chart you showed. Have you noticed anything here?


EDIT:
Also, I am considering SLI GTX470's very soon. May depend on BF3 scaling. I am curious about your ruminations on that topic too.

 

[AdamK47]

If you have a Fermi GPU and really want to test stability of your GPU, run Folding@Home projects 7620 and 7621. I can run OCCT all day long with 3-way SLI and get 100&#37; stability at 850MHz on all three GTX 580s. Run three 7620/7621 projects and the GPUs will be 94+ C. I've even had the protection circuitry tripped on my Corsair AX1200 power supply when running these. If instability is detected, Folding@Home will report it and stop the execution. OCCT runs the GPUs at 5 degrees cooler or less with 500RPM to 1000RPM less fan speed and will never report instability for me at these clock speeds.

I gave IDontCare instructions to run it, but I think HeDontCare.

 

[AdamK47]

Good question, I'm actually quite ignorant of OC'ing in the >1 video-card space but I am glad you brought it up because I aim to SLI my GTX460's and I was assuming OCCT was SLI-aware.

I take it that OCCT is not SLI-aware?

OCCT does. It runs the GPU test as crysis.exe. That way it uses the Crysis SLI profile when running the test.

Edit: Or the Crysis Crossfire profile.

 

[AdamK47]

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.

• 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)

Sorry to nitpick, but I find 5 runs of LinX far too low. 20 is even a bit low in my opinion. I've always run LinX on a timed basis. From between 8 and 12 hours. I've had many overclocks fail after 3 to 5 hours. I also run with all logical cores since this generates maximum usage and heat. It allows me to detect errors earlier.

I agree with Prime95 though. Been using it for a decade or more.

 

[Idontcare]

Awesome! I will do this after I install the new cooler. I'll make a public google doc and we can all add our data, and we can see a line for each persons CPU. I am curious about how mine lines up against others here, and also how wide the standard deviation is.

Do you know, in your rapid method, for each time you lower the Vcc, do you know if the temperature affects AI Suite's CPU speed that it settles on? I am trying to figure out what differences in CPU cooling/temperature would have on that chart you showed. Have you noticed anything here?


EDIT:
Also, I am considering SLI GTX470's very soon. May depend on BF3 scaling. I am curious about your ruminations on that topic too.

I realized when I saw this graph in your post that I failed to detail the relevant test conditions

I will update the graph, and knowing photobucket the pic will auto-update as well once our browser cache clears out.

In the meantime, what you are looking at there is the minimum Vcc required to fully pass 5 cycles of IBT (roughly 1hour per datum point) for every multiplier between 16x and 50x.

It took me roughly four days to generate the data. I'm still wrapping up the relevant data analyses, but here is some more "data dump" to appease your curiosity

Here's the "peak" core temperature experienced during the 5 runs of IBT:

And here is the peak power-consumption (measured at the wall, includes whole system but not the LCD):

 

[gevorg]

Awesome temperature and power graphs!! I hope Ivy Bridge would make 5GHz as easy as 4GHz in Sandy Bridge.

 

[lehtv]

Looks like 4GHz is the sweet spot for MHz/Watt when overclocking. After that it starts to climb noticeably more steeply... If IB is anything like SB, I'll probably OC mine to 4GHz and cool it passively 8)

 

[MrTransistorm]

Looks like 4GHz is the sweet spot for MHz/Watt when overclocking. After that it starts to climb noticeably more steeply... If IB is anything like SB, I'll probably OC mine to 4GHz and cool it passively 8)

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.

 

[Idontcare]

Sorry to nitpick, but I find 5 runs of LinX far too low. 20 is even a bit low in my opinion. I've always run LinX on a timed basis. From between 8 and 12 hours. I've had many overclocks fail after 3 to 5 hours. I also run with all logical cores since this generates maximum usage and heat. It allows me to detect errors earlier.

I agree with Prime95 though. Been using it for a decade or more.

That's no nitpick. You highlight the reality of statistics and reliability analysis. Accelerated stress testing is all about placement on the Weibull distribution curve and backing out the confidence level from there. Sounds like you prefer to operate farther to the right on the curve, higher confidence intervals, no issue there.

If 5 passes on IBT is to be considered the "bronze standard" and 20 passes the "gold standard" then we'll reserve AdamK47's stress testing thresholds for the "platinum standard" level

Now if we can gain a some manner of consensus that more people really do stress test for 8-12 hrs on LinX then I will happily adjust the thresholds in the OP, the guide is not meant to be written in stone.

Also...regarding your logical core vs physical core statement, there must be something else afoot here because I experience the exact opposite in LinX/IBT

Higher temps with 4 threads, also higher power-consumption at the wall as well as higher gflops.

 

[Idontcare]

Looks like 4GHz is the sweet spot for MHz/Watt when overclocking. After that it starts to climb noticeably more steeply... If IB is anything like SB, I'll probably OC mine to 4GHz and cool it passively 8)

Good eye! :thumbsup: The peak is exactly at 4GHz, for my CPU and setup at least

 

[NoobyDoo]

since hci memtest is run from windows, a large chunk of memory will not be tested, right ?

if your discrete/on-board gpu uses system memory, that too will not be tested, i think.

 

[Idontcare]

since hci memtest is run from windows, a large chunk of memory will not be tested, right ?

if your discrete/on-board gpu uses system memory, that too will not be tested, i think.

I don't really profess to have a technical working understanding of the limitations of HCI memtest so I can't really answer that.

What I can say is that this line of thought is exactly why I avoided using HCI memtest when I made a thread about the preferred memtest program in the Memory & Storage sub-forum.

I too was under the assumption that HCI in a windows environment would be inferior to memtest86+.

Then I had a problem where IBT/LinX would fail (error detected) even when my CPU was at stock. So I ran memtest86+ for >12hrs, some 12 passes, and it had zero errors detected.

Then I ran HCI, didn't even make it through one pass (<100% coverage) and bam, errors galore were detected.

I then downclocked my supposedly stable (per memtest86+) ram and reran HCI - ram came out stable. Then I reran IBT/LinX and it passed as well.

Conclusion - HCI works when memtest doesn't.

One thing about HCI is that you can specify in MB the amount of ram each instance is supposed to allocate. I just manually tell it to allocate all the physical ram. If you watch disk usage in Resource Monitor you can see a bunch of activity at first as the OS swaps out the cached stuff in ram to the disk, then the disk activity dies down while the memory in use goes to 99%.

I'm not saying don't use memtest86+, I love the unattended automated utility of a bootable thumb-drive, but I am saying that I have experienced first-hand that memtest86+ is not as aggressive/reliable of a memory test program as HCI for whatever reason.

 

[sangyup81]

I don't really profess to have a technical working understanding of the limitations of HCI memtest so I can't really answer that.

What I can say is that this line of thought is exactly why I avoided using HCI memtest when I made a thread about the preferred memtest program in the Memory & Storage sub-forum.

I too was under the assumption that HCI in a windows environment would be inferior to memtest86+.

Then I had a problem where IBT/LinX would fail (error detected) even when my CPU was at stock. So I ran memtest86+ for >12hrs, some 12 passes, and it had zero errors detected.

Then I ran HCI, didn't even make it through one pass (<100&#37; coverage) and bam, errors galore were detected.

I then downclocked my supposedly stable (per memtest86+) ram and reran HCI - ram came out stable. Then I reran IBT/LinX and it passed as well.

Conclusion - HCI works when memtest doesn't.

One thing about HCI is that you can specify in MB the amount of ram each instance is supposed to allocate. I just manually tell it to allocate all the physical ram. If you watch disk usage in Resource Monitor you can see a bunch of activity at first as the OS swaps out the cached stuff in ram to the disk, then the disk activity dies down while the memory in use goes to 99%.

I'm not saying don't use memtest86+, I love the unattended automated utility of a bootable thumb-drive, but I am saying that I have experienced first-hand that memtest86+ is not as aggressive/reliable of a memory test program as HCI for whatever reason.

When you used memtest86+, did you run loops of only Test 5? I seem to get satisfactory results this way. Satisfactory meaning if it passes 5 loops of memtest86+ test 5, Windows never crashes on me with all other things being stable. I'll try to see if I can get errors from HCI that I don't get from memtest86+ one day

 

[MadScientist]

HCI Memtest

According their manual the best way to run HCI memtest is:
How much RAM to test: To keep the test running smoothly and quickly, you should only test the amount of RAM that is unused, rather than the total amount of RAM in your system - otherwise your computer will spend 90% of the test reading and writing to your hard drive, rather than testing RAM. By default MemTest will check all RAM that is not in use by other applications. If you want more control, you can use the Windows Task Manager to determine how much RAM is free for checking. As a rule of thumb, Win9X uses about 32MB of RAM when nothing is open, and Windows2000/XP uses about 64MB. Vista uses quite a bit more, depending on what features you have enabled. Even though you cannot directly check used RAM, Windows dynamically moves the location of most of its subsystems, so most of your RAM will be checked eventually.

No Windows program can directly check the RAM used by the OS; this is a fundamental limitation of using a modern OS. If you need to check every byte, consider purchasing MemTest Deluxe, which boots off of CD for unfettered access to RAM.
http://hcidesign.com/memtest/manual.html

Their manual also states:
How many instances of MemTest: If you have a multi-core or multiprocessor machine you can make the memory test more effective by running multiple copies of MemTest at the same time. Start at least as many copies of MemTest as you have cores and then divide the amount of RAM to test between them equally.

Even though I have a 4 core CPU I have to run 8 copies, 1825MB each. I get this message if I try to divide my available ram, 14602MB, equally between 4 copies.

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.

 

[AdamK47]

That's no nitpick. You highlight the reality of statistics and reliability analysis. Accelerated stress testing is all about placement on the Weibull distribution curve and backing out the confidence level from there. Sounds like you prefer to operate farther to the right on the curve, higher confidence intervals, no issue there.

If 5 passes on IBT is to be considered the "bronze standard" and 20 passes the "gold standard" then we'll reserve AdamK47's stress testing thresholds for the "platinum standard" level

Now if we can gain a some manner of consensus that more people really do stress test for 8-12 hrs on LinX then I will happily adjust the thresholds in the OP, the guide is not meant to be written in stone.

Also...regarding your logical core vs physical core statement, there must be something else afoot here because I experience the exact opposite in LinX/IBT

Higher temps with 4 threads, also higher power-consumption at the wall as well as higher gflops.

Perhaps the Gulftown is different. I have always had higher temperatures with 12 and 11 threads. If you have time, would you mind testing yours with 7 threads?