Curious stress test results

[trents]

So I have this 6600k overclocked to 4.6 for 24/7 use.

It's rock stable in daily use including running a pretty demanding game like Project Cars.

I can pass 8 hr. of XTU stress test, 3 hr. of OCCT large data set, 4 hr. of Realbench and at least 2 hr. of Prime95 v. 27.7. All no sweat.

At the same vcore and other OC settings I instantly fail Prime95 v. 28.10. I mean it won't go one minute. Not even to the AVXII stuff yet. If I put all OC settings on auto at 4.6 it passes Prime95 v. 28.10 okay but at the cost of much higher voltages. This doesn't pose a temp problem for me because I have great cooling for an i5 but I find it odd that I fail the newer version of Prime95 so quickly when I can pass every other stress test known to man. I don't know what to make of it. Oh, it's the second core that always fails.

 

[EarthDog]

Weaker core.. give her more voltage.

 

[trents]

But why doesn't that same core fail in all those other stress tests?

And is there a way to give more vcore to just one core? Seems like I have been able to do that in the past at some point with older CPUs.

Another question I have is, with Skylake CPUs can you control the cache voltage separately from the core voltage? I can't seem to find that setting in mhy bios. I could do that on my socket 1151 ASRock/Haswell board. Present mobo is the one in sig.

 

[EarthDog]

Because of the version of p95.

I dont think you can give voltage to each core.

Cache is part of vcore in skylake...it was not in haswell.

 

[custom90gt]

I've heard 28.10 is much more difficult to pass - not sure why though. I say if it runs everything you want to run at the lower voltage, run it that way...

 

[demien88]

P95 is doing same thing to me after 5 minutes or 10 in it one same core always fail.... Tried many other stress test tools and it passed. Avoid using prime something is wrong with it

 

[mackerel]

There is nothing known wrong with newer Prime95. If your system fails, it is almost certainly unstable under that workload condition. If you are sure you do not need it to be Prime95 stable, you can ignore it. I do run tasks equivalent to P95, thus I tend to have to use lower overclocks, more voltage, and better cooling than others might.

 

[EarthDog]

P95 is doing same thing to me after 5 minutes or 10 in it one same core always fail.... Tried many other stress test tools and it passed. Avoid using prime something is wrong with it

There is nothing wrong with P95...

 

[demien88]

How to explain then IBT passing on extreme... AIDA64 passing for 2 hours Uniengine etc... We are also not alone with this issue I start to believe its some type of bug. I never had that issue on 960GM-VGS3 FX but wnen I changed to 990x then the problems started so it might be the board itself ****ing bricks. Tried googling not a single solution so I just skiped prime and used other software.

 

[EarthDog]

Again, not a bug. It is a more difficult stress test to pass than the others. Not sure of the details but you can't look at it from such a stance of 'it passed b, c , and d, so a must be flawed'.

 

[mackerel]

Prime95 is misunderstood and I keep meaning to write up a mythbusting guide to it some day... but in very short, it is designed to do a specific task as best it can and will try to extract all the performance it can out of a CPU while it is running. I run tasks like those, using the same math library as Prime95 but in other software. I can run them stably, but I know I can't overclock as far as others without needing more voltage and cooling to get there. So once again, if you don't ever do this type of task, you could arguably skip it and just remain stable at other things.

In theory, it is possible to be stable at Prime95, but not at something else! Each test will do what it does, and it is unwise to make generalisations if the other tasks are very different. Running and passing a variety of stress tests will give good confidence, but you should check any critical applications in particular. When I was doing non-K overclock for example, I found RealBench to crash more easily than Prime95 would, so I think that would also be a good stress test to add to the list.

 

[trents]

Yeah, it is in my list as you will see in the first post. I can run Realbench till the cows come home but on the same settings I can't get past the first iteration with later versions of Prime95 before worker #2 fails. That's the part that doesn't make sense. I would expect to be able to run Prime for at least a few minutes before errors if I can pass all those other tests for hours.

 

[mackerel]

If it is unstable, why expect it to run for any amount of time? If it was marginally unstable, maybe it could take some time, but to near-instantly fail signifies a clear instability.

 

[trents]

But don't you think passing extended stress testing trials with several other stress testing tools would indicate it is at least reasonably stable? I would not think this would put the overclock profile even close to "clear instability." That just doesn't jive. Either that or you are implying all those other tools are worthless.

 

[custom90gt]

But don't you think passing extended stress testing trials with several other stress testing tools would indicate it is at least reasonably stable? I would not think this would put the overclock profile even close to "clear instability." That just doesn't jive. Either that or you are implying all those other tools are worthless.

You're not the only one that's noticed the difference between v27 and v28 p95.

This article tiers 28.10 above 27.9: http://www.overclock.net/t/1621347/kaby-lake-overclocking-guide-with-statistics

Here is my thought to it, if you're stable in everything you need to do then run it that way. Prime95 27.9 is good enough for me (although I tested this go around with 28.10). Everyone is going to have an opinion on the best stability test, do what works for you. I often find that I can run Prime95 for 24 hours but Linpack fails after 4 or vice versa. I like to test with both of them for 2 hours when I'm trying to narrow down where I may be stable.

 

[mackerel]

But don't you think passing extended stress testing trials with several other stress testing tools would indicate it is at least reasonably stable? I would not think this would put the overclock profile even close to "clear instability." That just doesn't jive. Either that or you are implying all those other tools are worthless.

Each test will test different things. Prime95 will test things in certain ways others may not. If you fail Prime95, you are Prime95 unstable. It doesn't directly indicate you will be stable or unstable at other things. It doesn't matter at all if you are Aida64 stable, RealBench Stable, or whatever. They test different things. Many people do decide they don't need to be Prime95 stable if none of their workloads stress it in a similar way. That is ok. It is the argument about proving stability, which can't be done, and being "stable enough" for use cases as desired.

Let's turn it around, and test if Prime95 was faulty in some way. Restore your system to stock settings. Does that fail also? Even if it did, I'm willing to bet the system was unstable at stock before I believe there is a fault in Prime95, although of course either case would need investigated to determine the actual cause. Prime95 isn't just a stress test, it was used to find most of the biggest known prime numbers, and is still being used to find ever bigger ones. If there was any flaw in it, you can be sure there will be a lot of people who would be interested in fixing it fast. I can't say it is free of errors any more than that could be said of any software. I do recall they have found rare conditions in the past where it may provide bad results under specific conditions, and that has been identified and fixed. Also about a year ago, people started seeing errors under specific conditions, and questions started flying as to the cause. The result of that? A previously unknown bug in Skylake CPUs was the root cause.

 

[trents]

Yeah, it is in my list as you will see in the first post. I can run Realbench till the cows come home but on the same settings I can't get past the first iteration with later versions of Prime95 before worker #2 fails. That's the part that doesn't make sense. I would expect to be able to run Prime for at least a few minutes before errors if I can pass all those other tests for hours.

 

[EarthDog]

Perhaps none of those are using whatever instruction set is causing it to fail in P95.

 

[trents]

Perhaps none of those are using whatever instruction set is causing it to fail in P95.

I think that is probably the issue. That first instruction set iteration, whatever it is, trips that weak core right away unless voltages are jacked up significantly. I was able to pass 2 hr. of P95 28.10 this morning at 4.6 ghz with considerably higher voltages than it takes to keep the other cores stable running the same test. I'm doing some experimenting right now with P95 27.7 to see how much less voltage I can get by with to pass a 2 hr. test. Taking captures of HWMonitor to be able to compare he various voltage readings. CPU-z seems to agree with HWMonitor on my system concerning the voltages displayed in the current VID column.

I hate the way this Asus Z170 pro manages voltages. Seems to be unpredictable. The changes you make in bios seem to not to correspond well with what you actually get, if the software reporting can be trusted. I also have no control over LLC and that seems to have a mind of it's own and be very unpredictable. Changing the offset by x amount seems to trigger some unexpected change in LLC that results in a vcore value you didn't expect based on the change in offset amount alone. The other thing is you have to make .02 changes in offset voltage for it to show up as changed in software reporting. Making only a .01 change seems to really change nothing. I would have expected more settings and more granularity of control from a board in this class.

 

[EarthDog]

There are reasons why I strongly prefer manual voltage control over the adaptive stuff... really, its just plain easier. I don't care if it sits at 1.35V at idle...really makes no difference to me either way. If I want to save power, I shut down the PC. God Bless less than 20s to a fully active desktop.