Prime95 Blend Stability Issues?

EarthDog · Jul 26, 2016

Where are you seeing this TDP thing?

That is a good guess, and you may be right... who knows though... I would GUESS it doesn't have much to do with with it because it tends to peak early (within 2 hours) and then since it loops iterations, it will come around again to that same 'stressful' portion.

There isn't really a 'resolving' thing... if a worker drops, the others keep going.

Leegit · Jul 26, 2016

It was in HWMonitor. I guess TDP isn't technically the right term. Just mean the power draw of the CPU.

I had never seen anything over 50W (when stressing 4.8GHz at 1.35V which instacrashed) and then I woke up this morning with my maximum toggled at 80W but it was running 7 threads at 40W-ish... I figured it saturates so I waited a bit to make sure my temps were stable before I went to bed. Didn't want it creeping into the 90C region while I slept :-/

Leegit · Jul 27, 2016

I have some more information. It happened again and I noticed on the HWMonitor that the package total power usage went to 91W and the Uncore went up to 50Wish while its normally sitting at <1W.

https://en.wikipedia.org/wiki/Uncore

Is this normal? Anyone know?

wingman99 · Jul 27, 2016

Well I tested today to get the screenshot of the (rounding error) I was talking about and I could not duplicate it. The only thing that was changed is the memory 2400Mz from 3200Mz, I will get my new 3200MHz memory from RMA in about two weeks and I will try to test again so I can get a Screenshot.

Kenrou the prime version 28.9 default is not AVX, it is FMA3. When I tested this time I was paying attention.

Leegit · Aug 7, 2016

Been doing some more benching with Prime 95 Blend. Terrible results so far... I crash at 53 minutes when running at 4GHz even up to 1.35V (tried 1.20, 1.25, 1.30, 1.35) ... it seems to always crash right at 53m... something about that test.

wingman99 · Aug 7, 2016

If you pass large FFT and small FFT however blend fails the possibility is memory trouble or Prime95 at 53 minutes is detecting the possibility something went wrong with your CPU when you were overclocking to 4.8GHz. There is nothing wrong with prime95 I have run it 24 hours. Did prime 95 pass before you really tried overclocking? The only thing you could try with the Pime95 software is delete it and download fresh prime95 in case it is corrupted with all the testing.

Earth Dog I tested with the new 3200MHz memory From RMA to get the screenshot of the (rounding error) I was talking about and I could not duplicate it.

Kenrou · Aug 8, 2016

wingman99 said:
Earth Dog I tested with the new 3200MHz memory From RMA to get the screenshot of the (rounding error) I was talking about and I could not duplicate it.

This memory until a couple BIOS revisions ago (on my rig) needed 1.36v/1.38v to pass Prime95/Realbench as the mobo was undervolting it, even now that it works fine in XMP 1.35v i see it as 1.344v both in BIOS and HWInfo64.

EarthDog · Aug 8, 2016

wingman99 said:
If you pass large FFT and small FFT however blend fails the possibility is memory trouble or Prime95 at 53 minutes is detecting the possibility something went wrong with your CPU when you were overclocking to 4.8GHz. There is nothing wrong with prime95 I have run it 24 hours. Did prime 95 pass before you really tried overclocking? The only thing you could try with the Pime95 software is delete it and download fresh prime95 in case it is corrupted with all the testing.

Earth Dog I tested with the new 3200MHz memory From RMA to get the screenshot of the (rounding error) I was talking about and I could not duplicate it.

leave large fft alone. Stick with small and blend.

Figured as much. As the dev said, it's already enabled in the torture test. Perhaps you were running the project and not the torture test so that's why it popped up saying that as its not enabled by default in the project.

@ Kenrou - did you measure/confirm that value with a multimeter? As you likely know, software cn be off. That said, I've never had an issue with the slight vdrop like that before.

Kenrou · Aug 8, 2016

EarthDog said:
@ Kenrou - did you measure/confirm that value with a multimeter? As you likely know, software cn be off. That said, I've never had an issue with the slight vdrop like that before.

Nop (don't have one) but as i said since the start that i have had to bump voltage in XMP at stock or it would fail mem tests like blend, was completely fixed AFAIK around BIOS 1504. I knew i might have stability issues with this specific ram that had just come out a couple weeks back, but having to overvolt was a 1st to me...

Skylake has been an interesting experience to mess around with, was also around that version (possibly 1402) that the voltages needed for each overclocking step also dropped by ~0.08v compared to what i had at the start (0902). I made a long and confusing about it at the time i think in which i also started OC my ram.

EarthDog · Aug 8, 2016

They are like $20... good to have handy.

Leegit · Aug 8, 2016

I've completely characterized the required voltage from 3.5 GHz - 4.7 GHz (100MHz steps) for Prime 95 Small FFT. I have a bunch of graphs I can put up later when I get home. I gave up on Large FFT all together but I went back to Blend to try it out. The funny thing is I can get 4.5GHz to work on ~1.30V for over 2 hours. But if I try the default 4.0GHz, it just always crashes at 53m. I'm gonna try a fresh Prime 95 install. See if that helps at all...

wingman99 · Aug 8, 2016

EarthDog said:
leave large fft alone. Stick with small and blend.

Figured as much. As the dev said, it's already enabled in the torture test. Perhaps you were running the project and not the torture test so that's why it popped up saying that as its not enabled by default in the project.

@ Kenrou - did you measure/confirm that value with a multimeter? As you likely know, software cn be off. That said, I've never had an issue with the slight vdrop like that before.

No it was not the project you have to sign in for work that needs to be dun, I tried that I did not want to setup account. I think it was my new memory not working like it did before.

Blend in Prime95 rotates to large FFT.

Leegit · Aug 9, 2016

Prime 95 Small FFT Results ---> http://www.overclockers.com/forums/showthread.php/772650-OC-i7-6700K?p=7951976#post7951976

EarthDog · Aug 9, 2016

wingman99 said:
No it was not the project you have to sign in for work that needs to be dun, I tried that I did not want to setup account. I think it was my new memory not working like it did before.

Blend in Prime95 rotates to large FFT.

Well, not sure what you think you saw then. But, it's enabled by default according to the dev.

Also a worthwhile passage from the same guy...

Prime95's default test, Blend, is a cyclic workload for testing memory stability, and Large FFT's combines CPU and memory tests. As such, Blend and Large FFT's both have cyclic workloads which are unsuitable for CPU thermal testing."

wingman99 · Aug 10, 2016

EarthDog said:
Well, not sure what you think you saw then. But, it's enabled by default according to the dev.

Also a worthwhile passage from the same guy...

Thanks for all the replies. I was also wondering what the round off checking does? I have more errors when it is enabled.

Without putting too fine a point on it, it checks for roundoff errors- if you don't check for them, you might miss some miscalculations (fewer error messages).

The FFT is a bit of an approximation- imagine if multiplying two integers produced a decimal output, such as 20.31. Now, since the inputs were integers, you know the output should be an integer, so you round off the decimals and say the output is 20. But what if the output is 20.49? Is the error 0.49 and the correct answer 20, or is the error 0.51 and the correct answer is 21?

Prime95 generally chooses an FFT such that errors stay under 0.40, to 'eliminate' this sort of uncertainty. If a roundoff error is over 0.40, either the program should use a slower but more accurate calculation (a larger FFT), or the computer made a calculation error. Prime95 checks this by repeating the step that produced the roundoff error- if repeatable, use larger FFT. If not repeatable, hardware error! http://www.mersenneforum.org/showthread.php?p=439753#post439753

Some times you just need to ask the Qestion in the Prime95 forums since you don't believe me.

EarthDog · Aug 10, 2016

Man if there is anyone on the forum that can bork a quote and make their post confusing.................

What I took away from that is:

1. Earlier in the thread, you said AVX can be disabled on 7/22. Then on 7/23, you asked that forum how to disable it?????? Good stuff.

2. You didn't explicitly state if it was for the stress test or the application (as we know what round off checking does...). Nor did you reference where the P95 dev said its already enabled. I would link what I posted and see what else you can get from them.

http://www.mersenneforum.org/showthread.php?t=13808

mod @ P95 forum said:
I am pretty certain that it is always on in the torture tests. It's aim is to flag pcs(and the tests they do) that won't pass a torture test I think.

P95 Dev said:
It is.

Note, the mod is the same guy answering your question in your link.

http://www.mersenneforum.org/showthread.php?t=17718

Guy asking a question said:
I searched but couldn't find a good answer to this question. Do the Round Off Checking and Sum (Inputs) Error Checking options in the Advanced menu have any affect on the torture test and overclock stress testing? Should they be enabled when using just that functionality of Prime95? Seems like a lot of speculation about it, but no hard answers.

P95 Dev said:
Both are automatically enabled during the torture test.

In the end, it seems like it is doing additional checks when enabled?? Seems like overkill from default to me, honestly. Particularly considering it is already enabled in some form per the creator of P95 himself. He doesn't seem to think its needed in the torture tests. Now, maybe I can see it needed for a production math PC or data cruncher, but, the stress test really seems to do plenty for what most users use it for here. I don't remotely see a need to enable that unless they come back with something a bit more convincing than that. I'm all ears!

wingman99 · Aug 10, 2016

Well like I said before from my testing of round off checking finds a error faster so you don't have to wait as long, also if you want 100% stable system Prime95 won't miss a error. With the modern CPU performance why not enable I don't see Prime95 slowing down for me.

Kenrou · Aug 10, 2016

N00b question here, the error testing seem to be a threshold/percentage or am i reading it wrong ? it doesn't stop at the 1st error it finds, he passes over a certain quantity before it decides that its unstable or switches testing methods ? 1+1=2 but 1.99999999* is still considered valid at least for a couple tries ?

That being the case one would never have 100% stability no matter how many hours you put it through, there would always be a certain amount of errors that it skips through ?

EarthDog · Aug 10, 2016

wingman99 said:
Well like I said before from my testing of round off checking finds a error faster so you don't have to wait as long, also if you want 100% stable system Prime95 won't miss a error. With the modern CPU performance why not enable I don't see Prime95 slowing down for me.

Ive been 100% stable when using Prime without checking anything extra off for years. I'm good with leaving out the extra step. I dont 'science/math' my pc, and for all my uses over time, from email and browsing to F@H on the CPU to rendering and encoding, Its never steered me wrong... specifically going Small FFT, then blend. If people want to use large, go for it. But for 99% of people, even enthusiasts here, default settings are plenty.

I used to be an '8 hour for each (small and blend) kind of guy, then years ago moved to 4 hours for each, and haven't looked back...... that is, until I ditched P95 and went to AIDA.

Leegit · Aug 11, 2016

EarthDog said:
Ive been 100% stable when using Prime without checking anything extra off for years. ... Its never steered me wrong... specifically going Small FFT, then blend ...
I used to be an '8 hour for each (small and blend) kind of guy, then years ago moved to 4 hours for each, and haven't looked back...... that is, until I ditched P95 and went to AIDA.

I can have an OC stable for 8 hours on small but then it will fail out in blend after 53m when it cycles... I thought AIDA was just for testing memory? Does it test the CPU as well? Or are you using Prime95 Small and AIDA in conjunction??

Prime95 Blend Stability Issues?

Gulper Nozzle Co-Owner

Member

Member

Member

Member

Member

Member

Gulper Nozzle Co-Owner

Member

Gulper Nozzle Co-Owner

Member

Member

Member

Gulper Nozzle Co-Owner

Member

Gulper Nozzle Co-Owner

Member

Member

Gulper Nozzle Co-Owner

Member

Similar threads