Prime95 30.8 (big P-1 changes, see post #551)

[Prime95]

Originally Posted by Kaboom

Do you mean just having the option set, even though no listed programs are running?

Yes

I think the "problem", here, is that, with the same configuration, 30.7 used to get CERT work, while 30.8 does not.

This could well be a new condition in 30.8. I believe it was James that had some case where PauseWhileRunning and CERTs did not behave well. I don't remember the details.

[James Heinrich]

Originally Posted by Prime95

I believe it was James that had some case where PauseWhileRunning and CERTs did not behave well. I don't remember the details.

I don't think it was me. LOBES and Kaboom have reported such issues, but not me. I am a fan of PauseWhileRunning but the machine where I make use of that doesn't normally do CERT work anyways.
I don't think disabling CERT work just because PauseWhileRunning exists in the config is necessarily a good idea -- the program(s) listed there might cause interruption very rarely. Isn't there some kind of running performance average that Prime95 records recent performance vs expected performance that might serve as a better indication of reliable availability?

[James Heinrich]

Originally Posted by Prime95

30.8 build 15

This is listed as the preferred download version on https://www.mersenne.org/download/ but the homepage https://www.mersenne.org/ still shows "Prime95 version 30.7 released" as the top entry.

[Kaboom]

Originally Posted by Prime95

Yes



This could well be a new condition in 30.8. I believe it was James that had some case where PauseWhileRunning and CERTs did not behave well. I don't remember the details.

That's it, then. I have PauseWhileRunning active to stop running whenever I have compilation jobs or a VM running.

[Kaboom]

Originally Posted by James Heinrich

I don't think disabling CERT work just because PauseWhileRunning exists in the config is necessarily a good idea -- the program(s) listed there might cause interruption very rarely.

I second this opinion...

[Prime95]

Originally Posted by James Heinrich

I don't think it was me.

The change was made Feb 5th. I don't know if the bug report was posted on the forum or made via email/PM.

[undervoltTryhar]

Whenever I try this:



https://i.imgur.com/HJvs036.png






This happens:


https://i.imgur.com/KBUg0CS.png




Happened the same with latest 30.8, both fresh download.


Why cant I use fixed 122k FTT in place ? Should I use other values for vcore undervolt testing ?

[Batalov]

Originally Posted by undervoltTryhar

Why cant I use fixed 122k FTT in place ?

Because such FFT doesn't exist.
FFT are implemented for a certain list of sizes. In vicinity of "122" there are: 120, 128, 144 and so on.

Longer answer: "still why not 122?"
Because FFT is an algorithm that relies on nested iterations of size that is a small number: 2, 3, 5, 7 (at this time), and no more than two of the 3s, 5s, or 7s. A large number of multiples 2s is admissible.
122 = 2 * 61, and 61 makes it impossible until someone (for example you!) writes the code to implement a pass of exactly 61.

[kriesel]

And to add to what Batalov already posted, least-smooth (having larger largest prime factor) fft length tend to be less efficient, while power of 2 fft lengths are often very efficient relative to the near neighbors.
See for mprime/prime95, last attachment of https://www.mersenneforum.org/showpo...19&postcount=5 (right column, higher is better); for Mlucas https://www.mersenneforum.org/showpo...5&postcount=12 attachment, right colum, smaller is better; for gpuowl, http://www.mersenneforum.org/showpost.php?p=488535&postcount=2 last attachment; or for cllucas https://www.mersenneforum.org/showpo...71&postcount=2 last attachment (which shows power of 2 fft lengths 4 Mi and higher distinctly faster per k of fft length than other smoothness values.
Lists of 7-smooth fft lengths here.

[Greene Peng]

Prime95 app auto close during running and show 240K FFT size failure. can someone help advise the possible failure reason?
FATAL ERROR: Rounding was 0.49999143, expected less than 0.4
Hardware failure detected running 240K FFT size, consult stress.txt file.
FATAL ERROR: Rounding was 0.5, expected less than 0.4
Hardware failure detected running 240K FFT size, consult stress.txt file.


Info:
Intel Fishhawk Fall platform.
Prime95 (v30.8b15)

[Prime95]

Originally Posted by Greene Peng

Prime95 app auto close during running and show 240K FFT size failure.
Intel Fishhawk Fall platform.
Prime95 (v30.8b15)

These are typical stress test error messages that simply mean "something went wrong". It could be a FPU error, memory error, cache error, overheating or power supply issue. That is, the cause could be just about anything.

You can try stress testing fewer cores, less memory, or disabling AVX-512 / FMA / etc. to perhaps narrow down the problem.

I'm not familiar with "Fishhawk Fall". If this is a not-yet-released product, you can email me at woltman@alum.mit.edu to discuss if new CPU features could possibly be an issue. Yes, I understand you may be very limited in what you can discuss.