Question Prime95 large FFT instability

[torpeym]

Hi, yet another p95 failure question. I have a new build using an 8700k at stock settings (with Enhanced Turbo disabled) on a msi Meg ace mb, 32 gb (2x16) vengeance lpx 3200 with xmp enabled, with a noctua d15 cooler. PSU is seasonic focus plus platinum 750. GPU is an older msi 1060. Everything new except the GPU. System passes every test except p95 blend or large FFTs. Version is 29.8b3. The same result occurs with xmp or at the default spd settings. I have even tried downclocking the cpu, turning off turbo entirely, etc.

I’ve run memtest86 many times, usually using sequential CPU enabled, and all but twice it passed the tests with zero errors. The times it failed it found four errors, at two addresses, which were consistent between the two tests. What is weird is I have run this same test many times since those errors and it always passes now.

When p95 fails, it’s a round off error, and only 1 of the 12 workers actually stops - the other 11 happily continue indefinitely just fine. Temps usually in the high 50s, and never get above mid 70s for any of the p95 torture tests. The fails are on random workers and not at consistently times or consistent tests.

It does pass a large FFT if I limit the amount of memory to 25000 instead of the default it chooses - that test has been running for a couple hours now just fine.

Before I start trying to RMA components, is this more indicative of a ram, cpu or mb issue? What else can I try to narrow down?

If it matters I have done some OC experiments on this machine but nothing excessive, and never witnessed any issues (other than p95 fails).

This is a complete part list: https://pcpartpicker.com/list/JhvxD2

Thanks!

 

[alceryes]

Does HWiNFO64 give you any CPU 'performance limit reasons'?

 

[torpeym]

Does HWiNFO64 give you any CPU 'performance limit reasons'?

The only YEs is IA: Max Turbo Limit.

There doesn’t appear to be any throttling going on either, as the ratio is constant at 43 the whole test.

So, last night I was experimenting with custom large (146 to 8192) FFT using different amounts of RAM, increasing that to see when it failed. It never did, even when I bumped RAM back up to the max (29698). So I reran a blend test, which ran for 6 hours with no error (always failed within 90 minutes). Huh, the only thing I could think of that might have fixed it was I had previously reloaded the Default profile from intel extreme tuning. However, I rebooted and reran a blend test and it once again failed. I then reran intel extreme tuning and tried to reload Default, but it wouldn’t let me apply because it thought there was no change. So I made a change and then loaded default so I could apply it. Reran blend. Failed again after an hour. I wonder why all those tests passed until I rebooted?

 

[mdd1963]

Drop to a single stick of RAM, and test with each alternately....

I'd also research what timings are recommended/tested/approved for Vengeance LPX 3200 MHz at whatever RAM voltage, as certainly not all '3200 MHz' is equal...(some might require less stringent timings to the point where 2933 or 3000 MHz might actually be better anyway)

 

[torpeym]

Drop to a single stick of RAM, and test with each alternately....

I'd also research what timings are recommended/tested/approved for Vengeance LPX 3200 MHz at whatever RAM voltage, as certainly not all '3200 MHz' is equal...(some might require less stringent timings to the point where 2933 or 3000 MHz might actually be better anyway)
[/

Hmm, Friday night, I ran realbench stress tests for several hours, and all those tests passed, but now in HWiNFO64, there is an additional "performance limit reason" (IA: Turbo Attenuation (MCT)). What does that mean? I have not been able to find much googling for it. If i turn off turbo in the bios, this goes back to No, but when I turned it back on its back to Yes. Power cycling didn't clear that flag either. Other than that field, the machine seems to be behaving the same as it was before.

 

[mdd1963]

There would be no reason to disable standard turbo features...; as you paid for 4.7 GHz, I'd darn sure want it.

I'm not sure if 'Enhanced Turbo' of which you reference is a veiled renaming of MCE function or not, or, just standard Turbo...

If all cores go to 4.7 GHz under load, the MCE feature is enabled, with whatever name your MB maker is calling that feature.

In any event, Prime 95 blended mode is more RAM intensive...

Retest with a single stick, testing each...

I'd default your BIOS (so that standard turbo is enabled, which should allow a single core to hit 4.7 GHz, as I doubt that single core going to normal speeds is causing any Prime95 issues

Retest for Prime95 stability at stock CPU settings, with lesser RAM clocks (try 2666 MHz).

If your MB supports MCE, you can use it, but, it might require a VCORE bump, and, the extra 300-400 MHz across all 6 cores will raise temps of course...

I thought we were working P95 faulting out issues....?

If you only want gaming stability and/or Realbench, and you are currently happily satisfied, that's your call...

 

[torpeym]

There would be no reason to disable standard turbo features...; as you paid for 4.7 GHz, I'd darn sure want it.

I'm not sure if 'Enhanced Turbo' of which you reference is a veiled renaming of MCE function or not, or, just standard Turbo... If all cores go to 4.7 GHz under load, the MCE feature is enabled, with whatever name your MB maker is calling that feature.

Thanks for the help BTW! My motherboard (MSI) calls it Enhanced Turbo, and that is DISABLED. The usual behavior is that most of the cores hover around 4.3Ghz after booting, and then bounce around depending upon system load (as low as 800Mhz, and as high as 4.5 or so). I rarely ever see it go all the way up to 4.7Ghz. I am sure that MCE is disabled. One oddity I notice -- I almost never see it go up to 4.7.

In any event, Prime 95 blended mode is more RAM intensive... Retest with a single stick, testing each... I'd default your BIOS (so that standard turbo is enabled, which should allow a single core to hit 4.7 GHz, as I doubt that single core going to normal speeds is causing any Prime95 issues

I will do this over the next couple days.. I did do this a week or so ago, and I believe that each stick passed overnight runs of memtest86 AND did not cause Prime95 to fail out -- but I will retest to be sure nothing has changed.

Retest for Prime95 stability at stock CPU settings, with lesser RAM clocks (try 2666 MHz).

As I said in the OP, P95 fails out even at the default non-XMP settings (2133 MHz). This is with latest BIOS, and all bios settings stock -- with the exception of "Enhanced Turbo" with I disabled (it defaults to "Auto", whatever that means).

I thought we were working P95 faulting out issues....?
If you only want gaming stability and/or Realbench, and you are currently happily satisfied, that's your call...

Yes, I was using p95 to test whether all my new components are functioning correctly, as I would like to play around with OC. I tried realbench to see what results I got from it, since I can't make heads or tails out of why p95 is failing.

It was only after running realbench that the new CPU Performance Limit Reason / flag appeared (IA: Turbo Attenuation (MCT)) -- prior to that, the only Yes in the HWiNFO86 list was for IA: Max Turbo Limit.

Thanks!

 

[TechyInAZ]

Are you using Intel XTU for profile adjustments? If so why?