Consistent Error. Potential Bug?

[Shat]

Setup
CPU – Ryzen 5 3600x
Cooler – Noctua U9S push+pull
Motherboard – MSI B450i Gaming Plus AC (latest beta bios, issue occur on both latest non-beta and beta bios. Contacting AMD they told me to update to beta...)
Ram – Corsair 32GB 3200 C16 (CMK32GX4M2B3200C16)
PSU – Corsair SF600 Platinum
GFX – Gigabyte 1080 Turbo OC

Issue

I started observing BSOD and random reboot overnight sometime. (not heavy load, usually just downloading and/or automated stuff)


Diagnostic

This led me down the rabbit hole to diagnose the issue to see if I have a hardware problem.


I defaulted everything in BIOS and started running a suite of tests.
AIDA64 looks solid for ~1hr
OCCT free 1hr CPU/RAM/PSU all passes
Memtest all passes
Linpack Xtreme 10GB setting ~1hr all pass


However, I can consistently reproduce a failure using Prime95 on Worker 3 or 4 (Core 2 for my Ryzen 3600x), and it either fails on immediate test startup, or 1hr ~40min into the test on FFT Length 896K. (This is using Blend stress test)


I dug around stress.txt and it mentioned that repeatable test might be a bug related to Prime95?
I'm running v30.3 build 6.


The annoying thing is, I sent my CPU/RAM/Mobo initially to retailer to have it checked out, and they couldn't reproduce the issue (I'm uncertain what version of Prime95 they tried or whether they had all the same default settings I had...)


I've ordered a new PSU and going to try that out, though I'm doubtful that's the issue as the PSU runs fine with an older Intel/DDR3 system I have.

[Uncwilly]

Originally Posted by Shat

I dug around stress.txt and it mentioned that repeatable test might be a bug related to Prime95?
I'm running v30.3 build 6.

That bug was addressed with the release of 30.6 build 4. So, start there with that.

Welcome to the forum. Hopefully we can help you. And you stick around and contribute to the project.

[Shat]

Thanks for the welcome!

I've just downloaded 30.6b4 and... upon starting stress test using Large FFT setting, it immediately gave me a
"FATAL ERROR: Rounding was 0.5, expected less than 0.4
Hardware failure detected running 480K FFT size, consult stress.txt file)"


This occurred twice in a row when I restarted the test...
I just restarted the test a third time and now it seems to be fine...


Always core 2 workers still...


Is there any logs I should collect that will help determine whether this is a bug?

[Viliam Furik]

Quick question: Did you download the 30.6b4 from the first post of the link that Uncwilly provided? If so, that was 30.5b2.

Version 30.6b4 is here, post #256 in the same thread.

[Shat]

Oh no, I went to the directory link (removed 30.5b2 file from the url) and found the latest version 30.6b4 on there.

It seems to have passed 896K test for now, but the initial 2 failure on Core 2 still worries me...

I'm unfamiliar with GIMP and Prime95 bugtracking/codebase, is there anywhere I can see a raised bug list or source code/bug fix detail etc?

[Uncwilly]

That thread is the place to track progress. The post that I linked to has the updates. Bugs used to be tracked as a separate post, but that hasn't been done for this version. The code is not on git or other locations. Every once in a while George shares it (I think once a version is stable).

[Shat]

Hmm, does blend test do random FFT length at random time?


I noticed my core 2 immediately failed (seems like it always does this on cold start) a couple times now and seeing result.txt, it's always 480K now on 30.6b4.

[drkirkby]

It seems highly likely to me you have a hardware fault, and I very much doubt it is the power supply. Although you have ordered a power supply, I don't think it is the PSU as I doubt it would only affect a particular core, but your problem is related on core #2. I would suspect the motherboard, RAM or CPU.

It would be worth running with one DIMM for a while. Performance would drop a bit, as you would only be using one of the two memory channels. Then swap the DIMMs over. See if the problem changes with DIMMs. If not, I suspect you are down to the CPU or motherboard.

It would be worth taking screenshots, photographs or log files showing the problem. You can then present these to the seller. Also show your BIOS to indicate you are not overclocking things, since overclocking always increases the chances of running into problems.

[paulunderwood]

Originally Posted by Shat

This led me down the rabbit hole to diagnose the issue to see if I have a hardware problem.

You are running at stock on good hardware.

My Intel Haswell will not run a Linux kernel unless the CPU is slightly overvolted. Maybe upping the CPU voltage a tad -- say 0.05v -- on your system will solve your problem.

Another thing that might be worth a try is underclocking the RAM.

[Shat]

Hmm, an update to this.


Got a new PSU and yea, I still observe Core2 worker failure.


I tried various setup with RAM as suggested, both RAM, 1 in DIMMA, 1 in DIMMB, then the other 1 in DIMMA and DIMMB.
They all came back with the same result as follows.


Nearly every time, it'll fail within 2min on blend with the following error on Worker 3 or 4

FATAL ERROR: Rounding was 2.352155227e+16, expected less than 0.4
Hardware failure detected running 480K FFT size, consult stress.txt file

Although every 1 out of 4 tries or so, it'll go through and run for a while, but in the end fail somewhere else with that error.


Is there a known bug around this? Or am I seeing actual hardware error.

[kruoli]

It is highly likely that this is a hardware problem. Since you "sometimes" experience different behavior when running the same test, we can be quite sure that the software is working as expected.