Overclocking and stability testing

[biostud]

There has been some discussion about stability testing of CPU's and the best way to do it. We already has this thread: http://forums.anandtech.com/showthread.php?t=2195063

So I was thinking about doing a survey/log over different stability tested systems. I was thinking about creating different categories of stress testing and time intervals depending on user input.

I know we all have different systems and different work patterns, so we are not going to get anything statistically significant, but we will at least get some kind of data.

The categories I was thinking are:

Maximum heat/CPU stress testing (prime95, IBT, LinX)

"Normal" stress testing (OCCT, ASUS realbench, 3Dmark etc.)

Other (various personal ways of testing)

For time intervals I'm not really sure, but would like some user input based on your ways of doing things.

I was think something like:

Less than 2h
Less than 8h
more than 8h

Then basically you should post your system settings and your way of having stress tested your system and I will put you in the correct category.

Then we also have to monitor crashes. I would suggest two categories. Hard crashes and soft crashes. Hard crashes are BSODs, unresponsive OS, spontaneous restarts/shutdowns. Soft crashes are programs crashing/freezing but otherwise a functional OS. I know that crashes be related to lots of other factors than CPU o/c but we really don't have other ways of measuring stability in real work scenarios. (I'm very much open to suggestions of better ways to do this)

So first make a post stating your method of stability testing (day zero), then whenever you any experience any sort of crash make a post here about what what you were doing (what software) and how many days has passed since day zero.

Maximum heat/CPU stress testing (prime95, IBT, LinX)

Less than 2h

Less than 8h

more than 8h



"Normal" stress testing (OCCT, ASUS realbench, 3Dmark etc.)

Less than 2h

Less than 8h

more than 8h


Other (various personal ways of testing)

Less than 2h

Less than 8h

more than 8h

 

[DrMrLordX]

Okay, I'll bite.

CPU: x2-220, unlocked to three cores (1, 3, 4), 3.6 ghz (stock vcore). NB is 2100 mhz. RAM is DDR3-1600 6-7-6-18 1T.

Maximum Heat
Prime 95(blend): More than 8 hours
Linpack (via LinX frontend): More than 8 hours

Other
Memtest86+: More than 8 hours

No reported crashes/failures yet under Win7 or Xubuntu 14.04.

 

[BonzaiDuck]

With hindsight, for anyone suddenly enamored of the overclocking obsession, I would advise as follows.

1) Keep your favored workstation in operation for a few months while you simply test the new OC'd system.

2) Once you have all the drivers, stress-test and monitoring software set up on the new machine (and before overclocking) -- clean up sources of red and yellow events in the event logs. Make sure your AV/firewall software is installed. Clone the boot/system disk and set the clone aside.

3) Take pains with the overclocking; extend your understanding of other components like RAM for a clearer view of desired settings.

In answer to the original question, I use Prime95 (all three tests) in short 2-hour runs to obtain initial stability. After that, maybe 30 runs of IBT (all threads -- "Maximum") or 50 iterations of affinitized LinX. I suppose this translates into the choice "less than eight hours," but -- cumulatively for a battery of tests -- "more than eight hours."

Upon passing, I will then bump up the load VCORE through appropriate settings by no more than 10 millivolts, and call it a day.

But there are other potential problems. You may pass your marathon stress-tests, but the system may experience occasional/intermittent crashes at what seems like idle operation. There is the puzzling Stop Code 124, likely memory or IMC-related 09C -- or more easily interpreted 101.

Sometimes the 124 code can result from limp "digi-VRM" settings. Or, you may need to adjust PLL Voltage, or choose a new offset or LLC setting. Or -- your choice of RAM timings and command-rate (1) need extra voltage for VCCIO -- more likely related to code 09C.

These subtle problems can crop up some time after you'd decommissioned your previous workstation and installed your preferred software on the new machine. And all this being said, it is possible that a software installation can cause BSODs, unannounced resets and freezes -- while you've focused your attention on hardware and BIOS settings.

Used to be the case that I would run Prime95 for up to 14 hours. With the other tests, I now deem the shorter 3 to 4-hour tests perfectly adequate.

One more thing: keep a notebook of settings, length of test-runs and outcomes, temperatures -- and the range of voltage during stress program operation. Save the notebook. It will always be a good reference when or if you ever need to tweak the system further.

Also -- sad but true -- "Step 0" -- test and validate your RAM at the settings you hope will work. This is the most time-consuming, since 16 GB of DDR3 RAM can take several days to pass a "thorough" test. And I usually run those tests from a bootable CD, as opposed to running them in Windows.

 

[DrMrLordX]

I always run Memtest86+ from my bootloader. Not everyone has a Linux bootloader, though, so running it off CD/USB is just as good.

 

[BonzaiDuck]

I always from Memtest86+ from my bootloader. Not everyone has a Linux bootloader, though, so running it off CD/USB is just as good.

I could start another thread with these observations I'm about to make -- more recent experiences tweaking my old 2600K.

This was the first, and really only -- DDR3 system I'd built. I didn't pay a lot of attention to the XMP settings and profiles for these memory modules: I simply set the basic timings with a command-rate of 2. OF course, those items are declared in the memory spec sheet for my G.SKILLs.

So this year I re-examined "XMP." Anandtech had published an article urging people to use the BIOS feature, arguing that it would give "optimal" performance.

Then, I consulted G.SKILL tech support -- always willing to help. They confirmed my suspicion that XMP did somewhat more than simply make things easy for noobies: it provides more aggressive values for the so-called "Advanced timings" we usually leave set to "auto."

But then, I also discovered that setting all the timings and command-rate to "auto" would still give the spec timings (i.e., 9-9-9-24), but it would also show a command rate set to 1!!!

So -- that's nice -- I could feel comfortable thinking that the XMP profile didn't need any further tweaks.

But for a motherboard which has specs showing "1600(OC)" for the usual RAM people bought for it, a set of 1866-spec RAM would also be "OC."

I discover that the overclocked system (4.7 Ghz) with the XMP RAM profile and CMD=1 suddenly doesn't last very long in IBT or LinX.

Of course, the BSOD Stop Codes hold a clue: 09C. One then turns attention to the processor's IMC and the VCCIO voltage setting.

And, as you'd expect, a couple bumps in increase within the generally acknowledged 1.2V limit solves the problem.

Or -- I also find that I can increase VCORE to resolve the problem.

So the RAM testing is not only important initially. It's important that you choose the best settings before running those tests. We may be so used to putting in the basic timings and command rate manually, that we forget to follow the easy path and choose "XMP." And how would I know that XMP would set the command-rate to 1, unless I tried it?!

At this point, I think the VCCIO tweak becomes more necessary the higher the CPU speed. You might be able to show RAM running flawlessly at CMD = 1 and stock VCORE and speed -- with VCCIO on "auto." It looked to me that VCCIO had to be raised a little bit in proportion to successive overclock speeds.

 

[DrMrLordX]

Meh, I always do timings and subtimings by hand. It takes awhile, but as long as I have the system on test/tune duty I can afford to take my time with it.

So anyone else want to follow Biostud's template? Surely we have some more data points out there to add to the pile.

 

[Sable]

I had my system stable (i thought) at 4.8ghz. I'd also done runs at 5ghz.

The occasional crash occured though. Mainly standby/power down stuff. BUt the occasional freeze.

The coccasiojaly crash dropped stablillity and also somehow ruinined network port stability. windows was borked.

Unacceptable. SO 4.5ghz and rock solid after a completely new install of 7.

Now rock solid. Might push fior another ovcerlcock some time when someone make sa game wrtrhy of it. ;_)

 

[BonzaiDuck]

I had my system stable (i thought) at 4.8ghz. I'd also done runs at 5ghz.

The occasional crash occured though. Mainly standby/power down stuff. BUt the occasional freeze.

The coccasiojaly crash dropped stablillity and also somehow ruinined network port stability. windows was borked.

Unacceptable. SO 4.5ghz and rock solid after a completely new install of 7.

Now rock solid. Might push fior another ovcerlcock some time when someone make sa game wrtrhy of it. ;_)

ASUS?! ASUS P8Z77-V Pro?!! That's the second-generation 1155 model-line, and could easily have the same BIOS features. There is a "Phase Control" setting and a "VRM Duty Control" selection. I discovered that neither of these are going to be particularly stressed unless you change the Current Control setting.

So instead of the "limp-o" choices, I simply set them to "Extreme". You could take your chances with the "Manual" option, but would likely find yourself choosing "High."

4.6 Ghz is more comfortable for me than the sig-spec of 4.7, only because the load voltage at 4.7 is just under the Nehalem 32nm upper limit of ~1.375V. And I discovered that "games get away from me" with the RAM set to CMD=1 or "XMP-Auto." That is, my human reaction times can't quite keep up with the game. Put another way: I can't even get a "tie" with some score I'd previously set -- it seems that quick.

So I might run it a day or so at one setting and a week duration at the other. In my BIOS, under "Tools," I can save a total of eight BIOS configurations or "profiles" to select later. I've got "stock," "4600," "4700," "4500CMD1," "4600CMD1," 4700CMD1." I was working on "4700_BY_CORE" -- but it can wait . . . .

 

[BonzaiDuck]

Meh, I always do timings and subtimings by hand. It takes awhile, but as long as I have the system on test/tune duty I can afford to take my time with it.

So anyone else want to follow Biostud's template? Surely we have some more data points out there to add to the pile.

It was definitely my own inclination as well. I've discovered -- as a "senior" -- old habits are dying more slowly.

But the exchange with G.SKILL techs (as I said) confirmed a [slow-ly!] growing suspicion: the XMP settings "do more" than just make it easy for Noobs who complain on Egg reviews that their DDR3-1600's only run at 1333.

I'd wonder if ANYONE would tell me my VCCIO setting of ~ 1.13V is excessive: there's still about 0.07V of "headroom." That's what it seems to take to get 4.7Ghz stable with the stock RAM volts, timings and CMD=1 (the main factor for VCCIO increase -- but using the XMP profile and "Auto" for all the timings.

Otherwise, with a manually-set command rate of 2, my VCCIO only needs about ~1.09V.

Helluva difference, though!

Definitely -- WAnt more data points for BioStud's template. Me? I'm babble-mouth today. Too much coffee. I'll shut up for a while.