OCCT 4.4.2 or later -- Thoroughness or Reliability of the CPU:OCCT stress-test?

[BonzaiDuck]

What is your opinion about this?

I always use this test as a preliminary stability assessment tool.

The programmer(s) assert that running the test for 3.5 hours should assure there are no sources of instability to generate errors. The program has the ability to trap errors and end the test in an orderly way if the VCORE voltage is within a certain range of a certain and stable voltage.

After that, I may run IntelBurnTest or Prime95 or both. The Prime95 offers a short test which should discover errors within 15 minutes or an hour. But the thermal peaks of the OCCT test can be maybe 10C lower than the Prime95 options. The Small- and Large-FFT tests each have their value for testing. However, for making adjustments to an existing clock setting, these sorts of tests may not be necessary if one can run the others. One can choose either LinX or the OCCT:Linpack test, of which the first has a bigger guarantee of putting the burn on your CPU -- temperatures should be slightly higher.

For voltage adjustments to existing clocks after some hardware change (doubling your RAM modules, for instance), One might only need to run OCCT:CPU or LinX affinitized or both. But again, OCCT:CPU gives less thermal stress by some 10C degrees.

How much faith would you place in OCCT? For 3.5 hours? For 5 hours? Etc.?

 

[Fir]

I used to use it years ago.
Today it simply isn't enough.
I find that if a system can run the real bench 2.54 stress test with all available ram selected without crashing or showing "instability detected" for eight hours then it should run anything.

 

[BonzaiDuck]

Somehow, I had the impression that RealBench was a fairly mild test.

Also, remember that all of these programs get more or less regular updates.

But you understand what I mean or you should. I'm looking for shortcuts to re-validating my settings. To add detail to my original summary, I'd used RealBench as a preliminary test just as I'd used OCCT.

Without any further insights, I suppose I'll just have to run the whole battery of tests after filling my remaining RAM sockets. Ultimately for me, if LinX with AVX passes 25 iterations and GFLOPS stays within a 2-GFLOP range, it all proves out. I'd just prefer a test that's equal with milder thermal impact. And of course, I'd like the quicker test.

My experience has supported the idea that these re-validations of settings prove out with the 25-iteration LinX Max-problem-size and RAM if the GFLOP results are always or mostly consistent.

And since RealBench is a fairly mild test by my standard, I can probably burn up some time running it as well.

 

[Fir]

I like realbench because it's a very similar real world workload to what people do with their computers.
Other tests may make the processor hotter (AVX) but these aren't realistic expectations of what is done with the computer.
And if parts are overclocked, the additional strain put on the power delivery components (voltage regulators) can push them into unsafe territory since the factory cooling isn't designed for this. Some boards (X299) have trouble keeping up with AVX loads stock! Overclock/overvolt on these and you may get smoke. And that's never fun particularly if it wipes out a $2k cpu.

 

[BonzaiDuck]

I like realbench because it's a very similar real world workload to what people do with their computers.
Other tests may make the processor hotter (AVX) but these aren't realistic expectations of what is done with the computer.
And if parts are overclocked, the additional strain put on the power delivery components (voltage regulators) can push them into unsafe territory since the factory cooling isn't designed for this. Some boards (X299) have trouble keeping up with AVX loads stock! Overclock/overvolt on these and you may get smoke. And that's never fun particularly if it wipes out a $2k cpu.

I absolutely agree.

That's why, wanting to continue using LinX, I latched on to the GFLOPS variation idea. And for that reason, given today's weather stats and our house AC setting, I think 10 iterations can prove in conjunction with other tests, as opposed to 25. Peak Package Maximum shows 81C after an hour, while the core values bounce around in a distribution mostly in the 70 to 80C range.

"255, 256, 256, 256, 256 . . . " five iterations to go. . . .

Unless I'm misled, there are two things to test for a new build and overclock: the thermal performance, and the stability. Once you know the thermal results for a particular voltage neighborhood, only the latter matters.

As for the motherboard components, this Sabertooth Z170 board is an amazement to me. the VCORE VRMs never peg higher than 45C. You'd think a more extensive phase-power-design would be superior or that the board wouldn't perform as well as other top-end ASUS motherboards, but published benchies show that it does.

 

[TahoeDust]

2hrs of Realbench and I call it useable. 12hrs of Prime w/ AVX and 12hrs of Prime w/o AVX keeping temps 85c or less and I call it 24/7 workload stable. Those settings get documented, saved, and become my new default setting. Just got my 4.8GHz overclock there over the weekend....I celebrated.

 

[BonzaiDuck]

2hrs of Realbench and I call it useable. 12hrs of Prime w/ AVX and 12hrs of Prime w/o AVX keeping temps 85c or less and I call it 24/7 workload stable. Those settings get documented, saved, and become my new default setting. Just got my 4.8GHz overclock there over the weekend....I celebrated.

Well, you were right about RealBench.

The 10 iteration LinX run made it to the 7th or 8th, then logged an error. Yet, the GFLOPS had stayed on the money. I suspect that the problem is not with the VCORE, but could more likely be due to VCCIO or VDIMM. And we're basically talking about re-validating a 4.7Ghz setting. Beyond that, the temperatures are unacceptable to me, although the chip is binned for 4.8.

RealBench borked after maybe 20 minutes with an error. And of course, as we said, it runs much cooler.

So first I'll retest the RAM with HCI -- some 12 instances -- for maybe 200% coverage at 4.6, and see what stats on VCCIO HWINFO64 spits out so I have an idea of a starting point.

With this much RAM, I don't see an urgent need to get things back to 4.7 Ghz, but I'm sure it can be done.

I suppose since we're talking about over-stressing and over-heating, it makes sense to avoid lengthy tests. I don't run Prime95 for 24 hours anymore. I'll run some of those tests for 5 hours. If I have a "method," it seems to work with more variety in the testing and shorter tests. Also, oddly, there is one Prime95 test with manual settings that can show errors quickly with less than an hour's run.

AN UPDATE: This thread overlaps another on "Memory and Storage" about doubling the size of memory with an additional kit as opposed to a 2x16. It looks as though this sucker will run at 4.7 as I'd planned as a fallback -- without fixing a new VCCIO voltage: You just downclock the RAM a 100 Mhz or two. No . . . Prob . . . Ley-Mo."