how do i stress test the CPU to validate that it's stable?

[grandpa_boris]

i have a gigabyte 8PE667 Ultra 2 motherboard in my sonata case. the chip is P4 2.8 (sandra says 2.81GHz). i had recently lowered the core voltage to 1.45V (from recommended 1.55V). that dropped the CPU temperature to 36-37'C at low load (used to be 41-43'C @ case/motherboard temperature of 77'F/25'C). this is with a zalman 6500 and the (stock zalman 92mm) fan set @ 1430RPM.

i had used prime95 in stress-test mode to keep the CPU busy overnight, without any obvious problems. but it also didn't seem to really raise the temperature of the CPU too much (are there freeware/shareware tools to keep track of CPU temperature over time?)

i want to drop the core voltage lower yet and lower the fan voltage to drop its RPM down below 1300 RPM (perhaps going down to a simple 5V feed, zalman webpage implies it will work @ 5V).

short of waiting for the disaster to strike and the system to hang or crash, is there some way to proactively stress the CPU to validate its ability to work right at the lower voltage? i've seen references to people lowering core voltage to 1.1V.

i am not comfortable running sandra's burn-in test overnight unattended because i had to disable its "fail-safe" option -- it bailed out of the test immediately because it didn't like the low RPM fan. i am running it now in a tight CPU-only test loop, monitoring it as i type this. it got the CPU to 51'C after 10 iterations of the test. is this good? bad? neither? what temperature should i be looking for as a signal to me to shut off the test and crank up the fan?

(the core voltage had dropped to 1.42 volts during the test -- why? is this normal? expected?)

[dukla2000]

General answer: this is an art (not a science): depending on your apps your machine may appear 100% stable but with other apps it is not. For general purpose use my rule of thumb says 1 hour Prime is 90% confidence, 24 hours of Prime is 100% confidence (for me and my apps - browsing, folding etc). If you are running a life support system you may want a different thumb!

...(are there freeware/shareware tools to keep track of CPU temperature over time?)...

MBM has an ability to record a log.

...i am not comfortable running sandra's burn-in test overnight unattended ...

If you get MBM you can set some suitable alarms. (Must be a link on the links page here.) This starts trying to turn the art into a science: Sandra is good for thrashing all parts of the PC (e.g. including optical & HD drives), but arguably doesn't stress the CPU hard enough/long enough continuously. CPUBurn is among the heaviest CPU stresses: Folding is pretty good but a failure cannot be guaranteed a CPU problem. Prime is pretty good and any failure is 99.999% likely a hardware problem, but is only a CPU/memory test. (And there are better memory testers.)

(the core voltage had dropped to 1.42 volts during the test -- why? is this normal? expected?)

Normal & expected IMHO. Core voltage reports different from what you set are not in themselves cause for concern (because of the unknown accuracy of your sensor, BIOS & reporting s/w). Fluctuations are far more interesting: from 1.45 to 1.42 is normal (IMHO) - frequent fluctuation from a 'norm' of more than about 2% worry me. (MBM will report your long term average: +-2% from that for me is nothing to worry about). Most likely a sagging VCore is the power circuit genuinely struggling to cope with a sudden load increase: remember today's CPU can go from circa 5W power use at idle to 70W or more in a clock cycle, so there are serious challenges for the power supply. Also remember the 'power supply' includes the psu itself, as well as the mobo VCore regulation circuits. Even the best psu cannot make up for cheapo mobo regulation: in my experience my ASUS A7V333 was particulary poor for VCore regulation stability under fluctuating load. I have no opinion on Gigabyte.

ps - the temps you report (25 room, 35 CPU idle, 51 CPU stress) pass reasonable sanity tests. If anything the 51 may be 'underreported' with the CPU fan low. I would expect a 70W CPU to rise about 20C between idle and load (depends on the thermal efficiency of the heatsink/fan). 35 idle is hardest to sanitise: it depends precisely how idle, how good your case airflow is etc. Another useful temp is the air temp 1" from the CPU fan intake.

[Ralf Hutter]

IMHO, Prime95 is the best way to check for system stability. It compares the results that your system generates to a database of the correct answers and throws up a flag (shuts down with an error message) if the results are not what was expected. The other "burn in" and test apps (like Sandra and CPUBurn and BurnP6) just load your CPU, they don't actually test to make sure it's calculating correctly.

If you were running Prime95 and not getting much of a temp raise you should check in Task mamager to make sure it's actually getting 100% of the CPU time. Sometimes Prime95 will start in "Low Priority" and will appear to be running fine but will only be using a small percentage of the CPU time. This will not stress the CPU enough to give relevant results.

If it's not running at 100%, first stop the torture test (go to "File>stop test") then right-click on the "Prime95.exe" in Task manager and set it's priority to "normal". The go back to "Options>Torture Test" and start the test up again. It should now be using 100% of the CPU time. Let it run for 24hrs.

[fmah]

Why not run a burn program and Prime at the same time?

[halcyon]

For heat generation CPU Burn-In is even better than Prime95:

http://users.bigpond.net.au/cpuburn/

Hoever, Prime95 is (using Stress Testing) does better with iteratated calculation accuracy.

Sometimes esp. with overclocking or high temps, errors creep in that go unnoticed in most of the general use. You'll notice odd crashes of programs without any error message, but not necessarily any system stabilitty (i.e. OS does not crash).

Running Prime95 stress test for a couple of days in loop is usually a good way to find OC/temp/ram incompatibility problems that are hard to find.

I've even run memtest86 on a system for ONE WEEK (full test in a loop) without a single error at all, but after 5 minutes of Prime95 I had error/execution halted.

So, your best option is to use CPU Burn-in for heat generation and Prime95 for hard to spot stability problem debugging.

regards,
Halcyon

[grandpa_boris]

MBM has an ability to record a log.

i tried running MBM with this gigabyte m/b when i got it and it seemed like it was not supported. i hadn't tried downloading a newer version of MBM. i will try it. sandra's burn-in test will do a history graph, though i don't know how far back it can go. i should also probably twiddle the max/min parameters in that test to set the low CPU fan RPM threshold down to 0 and instead peg the test abort to exceeding some max CPU temperature. intel's spec sheet says max temperature is 71'C, so bailing out if the temperature exceeds 68-69'C is probably reasonable.

ps - the temps you report (25 room, 35 CPU idle, 51 CPU stress) pass reasonable sanity tests. If anything the 51 may be 'underreported' with the CPU fan low. I would expect a 70W CPU to rise about 20C between idle and load (depends on the thermal efficiency of the heatsink/fan). 35 idle is hardest to sanitise: it depends precisely how idle, how good your case airflow is etc. Another useful temp is the air temp 1" from the CPU fan intake.

idle is having a few web pages with minimal gif animation opened and having media player playing an MP3 stream from a webcaster as i type this .
airflow: the case has 120mm intake fan @ 5V, 120mm exhaust fan on fanmate (can't measure the voltage without doing something drastic to the wires, but i'd estimate it at 6-7V). the airflow isn't terrible, but i expect to be taking the box apart and apishly emulating ralf's cablegami and wire bundle sleeving to clean that up . the heatsink feels cold to touch when the CPU core temperature is showing as 35'C. i probably didn't let the burn-in run long enough to build up the temperature sufficiently. it seemed to stabilize at 51'C for a while, but the longer term heat build up may have pushed it higher. i'll need to run it for 24 hours making sure it's running at a higher priority as ralf suggested.

[MikeC]

grandpa_boris

You clearly answered what you mean by idle, but the more interesting question still hasn't been asked:

What do you do that really stresses the system?

The reason this question is interesting is that while some people's work/play requires every last processing cycle they can get, others never stress their systems at all. If you are a gamer or do computer-based music / audio processing, or PC videographer or anything time-critical that requires lots of crunching power, then the various detailed stress-testing methods suggested here work are perfect. Kudos to all the learned members

But if you never push the system, what's the point? Or if you never system that long and that hard? Just overengineer it anyway? This is the lawyers' and conservative engineers' standard worst case scenario strategy. Not always appropriate or necessary.

Case in point:

In my overclocking days , I'd push the CPU to the absolute limit and then back it off a notch and play games. Even with pretty loud high airflow cooling, I'd get the odd crash, but it didn't bother me much. Just games.

But the same one PC (back then) was also a work machine. The way I would test whether it was stable enough was to run the one operation that always let me know when the system was not quite perfect: Convert a large manual (tech manuals I wrote for a long time) from Framemaker or Pagemaker into a high res PDF file. This would often take several minutes. If things weren't quite right, the resulting file would open on the same computer but on any other, cause Acrobat on that system to crash.

That was my only serious stress/stability test. If the system passed that test, I knew everything else I did was hunky dory. Becauase I messed around with oc'ing and stress testing a lot in those days, I learned that even if my oc settings would not pass an extended CPU stress test without errors, it would still pass my PDF creation pass fine and run much faster than if I played it safe (ie, set it to pass the stress test).

Is there something you do that always stresses your system? That's probably your best realistic stress test right there

My 2 cents.

[Riffer]

I've used Hot CPU Tester:

http://www.7byte.com/

There is a free version on the site.

Quite good.

[grandpa_boris]

What do you do that really stresses the system?

here goes my geek quotent: the most stressful thing i do on my home system are 3D games

[grandpa_boris]

For heat generation CPU Burn-In is even better than Prime95:

http://users.bigpond.net.au/cpuburn/

Hoever, Prime95 is (using Stress Testing) does better with iteratated calculation accuracy.

i will download it tonight and give it a whirl. i'll also try other stress apps suggested by people in this thread.

the point of the exercise, in this case, is to make sure that the system remains stable as i keep lowering the core voltage to the CPU. i am operating on the theory that the lower the voltage, the lower the dissipated power until i hit the spot where the voltage goes too low to reliably switch gates. if the CPU stays much cooler at the lower voltage, i can crank down the fan over the CPU heatsink to be much quieter. unless i can run the chip really stone-cold, i will probably not turn the fan off. at 5V, it shouldn't be audible over the residual nose of the disk drive. so that's the primary goal.

secondary goal: i don't know what effect a thermal stress will have on the CPU's stability if the core voltage is lowered. so unless someone here can tell me, the only way to find out is to try stressing the CPU at lower voltage and see if it keeps stable within spec sheet's limits (under 70'C).

by the way, i haven't yet looked into the question of how the gigabyte motherboard handles the application of lowered voltage levels. i hope i can lower the voltage using gigabyte's easytune tool without storing the setting in the BIOS. that should allow me to descend at .05V increments, allowing the system a day of stress and stability testing before forcing the value in the BIOS settings. but i think it doesn't work that way and any change i make using easytune is applied immediately and is persistent accross system resets. going from 1.55 to 1.45 seemed to be immediately persistent, but i expected this to be a safe jump (and so far it seems it was). as i keep going lower it may become more hazardous.

[Zhentar]

which games? Some games stress the CPU more than others, but most can get pretty high CPU usage.

[grandpa_boris]

which games? Some games stress the CPU more than others, but most can get pretty high CPU usage.

NWN is one of them.

game stability is important, but i am really more concerned about the stability of the undervoted CPU in general.

[dukla2000]

idle is having a few web pages with minimal gif animation opened and having media player playing an MP3 stream from a webcaster ...

Then the 35C passes my sanity test for idle (folks with that mobo or more P4 experience can correct me as needed). Point is your CPU is active (i.e. not in something like the AMD STOP_GRANT mode) and to cause an argument about 20% loaded. So you would expect the temp to be measurably above ambient, say 5-10C above. (If you were STOP_GRANT idle the 1-2C above ambient is possible.)

Yup - methinks MikeC measured the Fanmate at minimum as 6V?

Looking back to your initial query: go for CPUBurn or similar to work out how high you can get your CPU temp. Prime is hard to beat for CPU stability: Ralf was more eloquent than me in highlighting it compares actual results against expected results which identifies specifically if you are unstable.

[grandpa_boris]

Why not run a burn program and Prime at the same time?

because it will leave the system completemy unusable for anything else . this is my primary workstation at home. i have to be able to at least telnet to work to read email.

[SometimesWarrior]

Yup - methinks MikeC measured the Fanmate at minimum as 6V?

Just to be picky: it's 5V. If 5V is at 6:00 on a clock, then 7V is 9:00 and 10.5V-11V is 11:00. Everything past that is just between 10.5V and 11V.

I marked off the voltages on four Fanmates, with varying fan loads. They're all the same, from what I can tell.

Okay, back on topic...

[grandpa_boris]

1 hour Prime is 90% confidence, 24 hours of Prime is 100% confidence

following ralf's suggestion, i have prime95 running at "above normal" priority for 2 hours now (monitoring it using the task manager's processes tool with "base priority" column displayed). in the first 5 minutes there was a brief spike in CPU temperature that took it up to 57'C, but for the last hour prime95 has been consuming 99% of the CPU and the CPU temperature hasn't exceeded 39'C. the motherboard temperature is 25'C.

MBM has an ability to record a log.

i tried the latest version of MBM - it doesn't support my motherboard.

[Ralf Hutter]

MBM has an ability to record a log.

i tried the latest version of MBM - it doesn't support my motherboard.

In your first post you stated that your mobo is a Gigabyte 8PE667 Ultra 2.

I went to the MBM site and did some poking around and it seems like your board is supported. Start at the main MBM page and go to the "Motherboards" link, then to the Gigabyte page. Scroll down and you'll see your board listed there.

The sensor settings you need are: "ITE8712F-1" = Case temp, "ITE8712F-3 Diode = CPU temp.

There's a thread about your board that discusses all the sensor settings here.

[Riffer]

MBM has an ability to record a log.

i tried the latest version of MBM - it doesn't support my motherboard.

In your first post you stated that your mobo is a Gigabyte 8PE667 Ultra 2.

I went to the MBM site and did some poking around and it seems like your board is supported. Start at the main MBM page and go to the "Motherboards" link, then to the Gigabyte page. Scroll down and you'll see your board listed there.

The sensor settings you need are: "ITE8712F-1" = Case temp, "ITE8712F-3 Diode = CPU temp.

There's a thread about your board that discusses all the sensor settings here.

I've done a lot of fiddling with MBM trying to get it to work with my Gigabyte GA-8PE800Pro, with little luck. So I can see grandpa_boris having the same problems with his 667.

MBM will read the CPU temp on boot, but then malfunctions and shows the temp at a constant 70C.

That having been said, Gigabyte has a utility that monitors the CPU temp, RPM's etc. I think it is called easytune4, but I am not at home, so not sure.

[grandpa_boris]

I went to the MBM site and did some poking around and it seems like your board is supported....The sensor settings you need are: "ITE8712F-1" = Case temp, "ITE8712F-3 Diode = CPU temp.

you are correct, the site says the m/b is supported, but the readings i get from MBM are so weird that it doesn't look like it's working well at all. this morning it was showing the case temperature at 64'F, while the ambient house temperature was 70'F. it also shows the CPU fan speed at 0 RPM which is obviously incorrect. i will play with the settings a bit more.

[grandpa_boris]

Gigabyte has a utility that monitors the CPU temp, RPM's etc. I think it is called easytune4

i've been using it to change the CPU voltage. i also have been using it to monitor the system, kind of. it gets easily confused so i don't trust it too much.

[grandpa_boris]

results so far:

1.4 volts isn't stable. prime95 fails.

1.425 volts is stable. i've been running prime95 for over 14 hours now. at about 8 hours it seemed to have stopped using up as much CPU as i was hoping it would, so i had restarted it with different parameters and set priority to "high". as i left the house, the ambient temperature was 70'F, the CPU was @ 32'C. the one oddity was that the Vcore was showing 1.34V and i've seen it dip to 1.2V. this is within 20% of the setting, but i don't understand the fluctuation. this is something like 10'C lower than the CPU temperature @ 1.55V (normalizing for the differences in the ambient temperature).

i don't think this chip is stable under 1.425. i will keep prime95 running for now -- not that i have much of a choice, being at work and it being at home . if it doesn't fail until i get home tonight, i'll use easytune to fix that voltage on bootup and start experimenting with slowing down the CPU fan, allowing the chip to get hotter by using cpu burn loads people recommended in this thread, and monitoring stability running prime95.

one comment: prime95 has to be running the stress test to set the priority right, it seems. it appears to deliberately set its priority to "low" at the start of the test otherwise, over-riding the previous priority setting.

[grandpa_boris]

prime95 had dashed my hopes for a cool-running, noiseless system. my system had generated errors at all voltages i tried so far. i am running a load now @ 1.475V. we'll see how long it lasts. at 1.40V the temperature reported by MBM was 36-37'C. i was entertaining ideas of slowing down the CPU fan, may be even turning it off. now, at the nominal 1.475V i am seeing 56'C at the same fan setting. MBM claims the voltage is actually 1.49-1.50V, but prime95 seems to not be spitting out any errors 3 hours into the run.

sigh....

[dukla2000]

- that's a shame.

But also comes back to the spectacular 'marketing' claims posted around the web - folk running at 1.2V or ambient plus 1 degree etc: OK it boots and they can browse but it is unknown stability! I figure if Prime is failing then I am worried what is happening in Quicken and my spreadsheets: if Prime is stable then most likely garbage out means I put garbage in!

[Ralf Hutter]

prime95 had dashed my hopes for a cool-running, noiseless system. my system had generated errors at all voltages i tried so far. i am running a load now @ 1.475V. we'll see how long it lasts. at 1.40V the temperature reported by MBM was 36-37'C. i was entertaining ideas of slowing down the CPU fan, may be even turning it off. now, at the nominal 1.475V i am seeing 56'C at the same fan setting. MBM claims the voltage is actually 1.49-1.50V, but prime95 seems to not be spitting out any errors 3 hours into the run.

sigh....

Sorry to hear that , but that's the reason to run Prime95. If you hadn't done it you would have been running along at 1.4V thinking everything was peachy when in fact it wasn't.

I'm not suprised at the temps you're getting with 1.475V either. the 800MHz P4's put out a lot of heat. I'm OK with my 2.4C's, but my 3.0C gets pretty toasty under full load at default Vcore. You're talking about CPUs that are putting out 75+ watts here, that's a lot of heat to get rid of.

Replacing your Zalman 6500 with a 7000, or a Thermalright SLK947 or SP94 might gets you a few degrees cooler, but not enough to really help things.

One more very important thing that might be the underlying cause of your MBM issues (and maybe even you general instability) is if you're running that "Easy Tune" utility alongside MBM5 at the same time. Both of these apps will be competing for the same sensors which will cause all kinds of fun things to happen, like voltages and temps to be mis-reported, or to be reading seemingly OK now but goofy in a few more minutes.

Make real sure that you're only running one of these at a time. Be extra careful to make sure that they're not running in the background where it's not very apparent, especially with Easytune. MBM's not too funky, if you turn it off, or remove it from the Startup floder it will not be running in the background, but I'm not too sure about Easytune.

[grandpa_boris]

after around 24 hours of prime95, i am going to consider 1.475V the lowest stable voltage this CPU will support. with the CPU fan set @ 1400 RPM and 100% load for hours, CPU hasn't gone over 51'C. it's a far cry from 37-39'C i was seeing with 1.4V, but it's the price i will have to pay for stability. for now, this system will remain as is.

next projects:

• a silent (really silent, no fans) printer server
  
  a silent, very low power 2-in/1-out firewall (probably based on http://www.soekris.com/ net4801 board)

[Ralf Hutter]

next projects:

• a silent (really silent, no fans) printer server

My router's got a built-in printer server and it's 100% passively cooled and dead silent. Can't you also get USB hub sort of print servers?

[grandpa_boris]

My router's got a built-in printer server and it's 100% passively cooled and dead silent. Can't you also get USB hub sort of print servers?

my initial plan was to use something like http://www.hawkingtech.com/prodSpec.php?ProdID=89. i have 2 printers that need to be hooked up (one general purpose, one photo), to be joined by a firewire/USB scanner and probably another printer (die-sub). i don't believe this kind of stand-alone print server supports multiple printers per unit. i could have a print server per printer (ignoring the scanner for the moment), but there is a more significant problem. i use photoshop (hence the photo printer). i described my network printing plans to my friends from adobe, and they told me that photoshop has a problem properly negotiating with a remote printer unless that printer is attached to a windows or mac server.

my plan is to buld a minimal-powered VIA system. EPIA-M i have now comes with 4 USB and 2 firewire ports and a PCI expansion slot, so i can get additional ports. i'll follow your suggestion with the samsung drive, which may allow the use of a morex 2699 case i am already familiar with or a smaller 2677R. or i'll get a 2.5" laptop drive: performance isn't a goal here, silence and stability are.

not to forget the scanner: my plan was to hook it up to the same server. ideally, i should be able to access it accross the network. i doubt that will work, though. so i am expecting to use VNC to run it remotely and store scans to disks shared from my other systems.

[dukla2000]

Damn - I followed my own advice and boy is it bitter!

6 weeks ago got a new KT600 mobo, plugged in my XP2100 that would run more or less OK at 13*166 or 16*133 on the old KT333. Fired up at 10*200 and ran for 6 weeks 'with no problems'. Famous last words: did a bit of Prime over the w/e and discovered no reasonable VCore would keep it running Prime for more than 3 hours at 200fsb! Same story with a DLTC XP1800 I have. So back to 166fsb - oh well, means the step to a Barton is getting closer

[Ralf Hutter]

Damn - I followed my own advice and boy is it bitter!

No matter what, I always run Prime95 and Memtest86 on new builds, even if I'm building a system on an actual Intel board.

My theory is that there's no sense building a system if the foundation is unstable. I've had bad hardware right out of the box and by running these stability tests it's saved me from a lot of troubleshooting down the road.

And for overclocking (or undervolting in our case ) it's even more important to do this because we're running our hardware out of spec.

[Seal]

as a watercooler+overclocker prime95 gets my vote!