Request for version 28.9 torture testing

[Prime95]

I'm trying to track down a possible problem with either prime95 version 28.9 or a specific unnamed CPU.

Please run a blend torture test for a few hours using these settings in local.txt:

CpuSupportsAVX=0
CpuSupportsAVX2=0
CpuSupportsFMA3=0

Surprisingly, I tried that on my quad-core Haswell box and one worker failed with a roundoff error after several hours. Happened twice. That machine has overclocked memory, so this is not conclusive. However, it is interesting since SSE2 FFTs should not be bandwidth limited and this machine has been stable for over a year.

IIRC, some AMD machines were reporting torture test issues with 28.9.

Please report successes and failures as well as CPU, OS, whether it is overclocked, which FFT size failed (if you can tell), or any other data that might be pertinent.

Thanks.

P.S. Don't forget to remove those local.txt lines before resuming testing!

[chalsall]

Originally Posted by Prime95

I'm trying to track down a possible problem with either prime95 version 28.9 or a specific unnamed CPU.

Surprisingly, I tried that on my quad-core Haswell box and one worker failed with a roundoff error after several hours. Happened twice. That machine has overclocked memory, so this is not conclusive. However, it is interesting since SSE2 FFTs should not be bandwidth limited and this machine has been stable for over a year.

IIRC, some AMD machines were reporting torture test issues with 28.9.

Please forgive me for this, but if you can reproduce the problem with your own kit that is much more valuable and expedient than asking others to try to help you.

Correct me if I am wrong.

[Prime95]

Gathering evidence is the best way forward here. My one datapoint says that one overclocked Haswell has troubles running a torture test - not very informative. Now if 5 of 10 Haswells have the same issue, that's interesting.

And if 5 of 10 Haswells, but 0 of 10 Skylakes, have the problem, that would indicate a possible Haswell hardware issue. But if 5 of 10 Haswells and 5 of 10 Skylakes have an issue, then maybe we are dealing with an something like an uninitialized variable in prime95.

I'd also appreciate the torture test run on a machine that actually does not support AVX. Such a machine would segfault if an AVX instruction was run, whereas my Haswell keeps chugging along happily.

[chalsall]

Originally Posted by Prime95

I'd also appreciate the torture test run on a machine that actually does not support AVX. Such a machine would segfault if an AVX instruction was run, whereas my Haswell keeps chugging along happily.

My sincere apologies. This makes sense.

I have grown up being constantly attacked.

It's not easy to work with. But what the hell, there are worse things to work with.

[kladner]

I have it started on an FX-9590, running at "stock". That is to say, it is a factory OC 8350, at 4.7 GHz. The benchmark section of Prime.txt is-

FullBench=0
BenchMultipleWorkers=1
BenchMultithreads=1
MinBenchFFT=512
MaxBenchFFT=4096
# BenchTime=30
OnlyBenchThroughput=0
OnlyBenchMaxCPUs=0
OnlyBench5678=0
BenchAllComplex=0

CpuSupportsAVX=0
CpuSupportsAVX2=0
CpuSupportsFMA3=0

That said, I have it running in the default "Custom" mode, from 8K to 4096K FFT. If you'd like a different setup, let me know. I could run it as a stock 8350, if desired.

[Prime95]

@kladner:

The CpuSupportsXXX entries belong in local.txt.

To verify these settings are correct, choose Options/CPU dialog. It should report CPU features along the lines of Prefetch, SSE, SSE2, SSE4.

[kladner]

Originally Posted by Prime95

@kladner:

The CpuSupportsXXX entries belong in local.txt.

To verify these settings are correct, choose Options/CPU dialog. It should report CPU features along the lines of Prefetch, SSE, SSE2, SSE4.

Ouch. This is corrected and restarted. Paying attention to details.....
EDIT: Double ouch. I have a test instance of P95, and a production instance. I was editing the wrong *.txt files.

[Mini-Geek]

I've got it running on my Haswell, I'll let you know how it goes...

[potonono]

Blend torture tests completed successfully with the extra lines in local.txt.

Windows64,Prime95,v28.9,build 2
Intel(R) Core(TM) i5-6600K CPU @ 3.50Ghz
CPU speed 2341.48 MHz, 4 cores
CPU features: Prefetchw, SSE, SSE2, SSE4
L1 cache size: 32KB
L2 cache size: 256 KB, L3 cache size: 6 MB

OS: Windows 7 Ultimate SP1 (64-bit)
Not overclocked.

[storm5510]

Originally Posted by Prime95

I'm trying to track down a possible problem with either prime95 version 28.9 or a specific unnamed CPU.

Please run a blend torture test for a few hours using these settings in local.txt:

CpuSupportsAVX=0
CpuSupportsAVX2=0
CpuSupportsFMA3=0

I have to test my new box when it get here. It's Ivy Bridge. I can try with the settings above and see if it works. I've been using the 64-bit version of Prime95.

[0PolarBearsHere]

Using a clean copy of 28.9 and those 3 lines, my 3930K machine threw two instances of

FATAL ERROR: Rounding was 0.5, expected less than 0.4
Hardware failure detected, consult stress.txt file.

Is there any flag to write a more detailed results.txt so I can work out where it failed? Many different FFTs have varying numbers of lines so I can't just go on a reduced number of successes.

Windows 8.1 PRO
32GB RAM