Prime95 Blend Stability Issues?

If you stress test using the actual prime program and not the stress test inside, then you would want round off enabled. However, just use the stress test. Round off checking is already enabled when running the stress tests!
http://www.mersenneforum.org/showthread.php?t=13808
http://www.mersenneforum.org/showthread.php?t=17718

Also, it doesn't enable additional tests. Its just for checking errors in the actual data when running their program. Otherwise, its already enabled in the stress tests (see links).

Kenrou said:

You're missing the point, if you pass 24h 26.6 (yes with floating point and rounding checks) you're for all intents and purposes 100% stable for anything except AVX. So if you only game and browse you will do it at the same settings with less voltage and therefore less heat. Whats the point of making the cooler/case fans work harder ?

If you're doing important work then yes i completely agree that you should use 28.9 with all the bells and whistles turned on. Just don't panic when you see the temperature shoot up

Click to expand...

Not every CPU fails overclocking in the same way that is why it is called overclocking, your taking every transistor and running it on it's edge and some flip 0 or 1. The 86 architecture is basic with added instruction sets like AVX it's all the same transistors it's not about AVX version 1 or 2. If you have a AVX instruction set that can't handle overclocking the transistor in that group bit flips that could happen when overclocking in your case. All the transistors in the CPU are made and run equal by intel at stock speed.

I will give a example when I run version 26.6 or 28.9 it takes the same voltage for me for 7 hours stability when I tested it. You did not have the same results we have different CPUs they look the same however we are talking about differences at the atom level.

Anyway prime95 26.6 checks 32bit not 64bit also it checks AVX version 1.0 not 2.0, I would suggest using the new 28.9 with 64bit unless your using a 32bit PC.

- - - Updated - - -

EarthDog said:

If you stress test using the actual prime program and not the stress test inside, then you would want round off enabled. However, just use the stress test. Round off checking is already enabled when running the stress tests!
http://www.mersenneforum.org/showthread.php?t=13808
http://www.mersenneforum.org/showthread.php?t=17718

Also, it doesn't enable additional tests. Its just for checking errors in the actual data when running their program. Otherwise, its already enabled in the stress tests (see links).

Click to expand...

If I don't enable it does not show the rounding errors during or at the end of the test..

Why did the developer say otherwise.. twice?

When a worker stops in the stress test, that is a rounding error, is it not? It shows instability... it doesnt show the type regardless. Worker stops = unstable.

wingman99 said:

Not every CPU fails overclocking in the same way that is why it is called overclocking, your taking every transistor and running it on it's edge and some flip 0 or 1. The 86 architecture is basic with added instruction sets like AVX it's all the same transistors it's not about AVX version 1 or 2. If you have a AVX instruction set that can't handle overclocking the transistor in that group bit flips that could happen when overclocking in your case. All the transistors in the CPU are made and run equal by intel at stock speed.

Anyway prime95 26.6 checks 32bit not 64bit also it checks AVX version 1.0 not 2.0, I would suggest using the new 28.9 with 64bit unless your using a 32bit PC.

Click to expand...

26.6 runs 64bit as well, and DOES NOT use AVX (starts at 27.7), thats why there's a few posts advising for it on Haswell against the newer versions.

Check the posts by CompuTronix (if there's another thread more updated pls let me know) :

http://www.tomshardware.co.uk/answers/id-2557021/prime95-test-question.html

http://www.tomshardware.co.uk/forum/id-1800828/intel-temperature-guide.html

EarthDog said:

Why did the developer say otherwise.. twice?

When a worker stops in the stress test, that is a rounding error, is it not? It shows instability... it doesnt show the type regardless. Worker stops = unstable.

Click to expand...

The rounding Error does not stop the worker it keep going it just shows Error while testing or when your done. I have tested back to back with rounding check off and on. With it off it wont show rounding errors. It actually says it while testing, it says rounding error.

Kenrou said:

26.6 runs 64bit as well, and DOES NOT use AVX (starts at 27.7), thats why there's a few posts advising for it on Haswell against the newer versions.

Check the posts by CompuTronix (if there's another thread more updated pls let me know) :

http://www.tomshardware.co.uk/answers/id-2557021/prime95-test-question.html

http://www.tomshardware.co.uk/forum/id-1800828/intel-temperature-guide.html

Click to expand...

Your looking at the wrong resource for information. what you need to look at is GIMPS-Free Prime95 site http://www.mersenne.org/download/

FreeBSD 32-bit 26.6 2011-04-08 Requires FreeBSD 7.0. May work on later versions, it is a 32bit prime95. The AVX features is in Version 27.1 for sandybridge

32-bit FFTs optimized for AVX capable computers. Intel Sandy Bridge computers should
see a 25% speed increase. http://www.mersenne.org/download/whatsnew.txt

Click to expand...

I don't know why you don't want people to test with AVX or 64bit? You can Manually disabling AVX testing however you can't turn it off in the CPU.

The 64bit version is from the website, the link is in the forums instead of main page (unless I'm completely mistaken on your meaning of 64bit).

http://www.mersenneforum.org/showthread.php?t=15504

Not a case of wanting, simply there is no point testing with anything above 26.6 if the person does not use AVX programs like I said in a previous post (also mentioned lots of argument over it which ended up happening again). I have tested it in 4 different computers, and so far yours is the abnormal here, 26.6 required less voltage on all of them to pass 24h comparing to 28.9, which would be the AVX instructions.

More voltage, more heat, more wear and tear, bigger electric bill on a daily basis for ZERO gain.


Edit : Out of curiosity, how do you disable AVX/FMA3 ? see no option for it in any version ?

wingman99 said:

The rounding Error does not stop the worker it keep going it just shows Error while testing or when your done. I have tested back to back with rounding check off and on. With it off it wont show rounding errors. It actually says it while testing, it says rounding error.


You can Manually disabling AVX testing however you can't turn it off in the CPU.

Click to expand...

if you have a rounding error/instability n p95 stress test, the worker will stop. Tbe rest of the workers keep going, yes.

Perhaps there is a difference in notification delivery (worker stopping or the whole test stopping? However, according to that dev i linked, twice, it's already on.

How do you turn off AVX testing in p95?

Kenrou said:

The 64bit version is from the website, the link is in the forums instead of main page (unless I'm completely mistaken on your meaning of 64bit).

http://www.mersenneforum.org/showthread.php?t=15504

Not a case of wanting, simply there is no point testing with anything above 26.6 if the person does not use AVX programs like I said in a previous post (also mentioned lots of argument over it which ended up happening again). I have tested it in 4 different computers, and so far yours is the abnormal here, 26.6 required less voltage on all of them to pass 24h comparing to 28.9, which would be the AVX instructions.

More voltage, more heat, more wear and tear, bigger electric bill on a daily basis for ZERO gain.


Edit : Out of curiosity, how do you disable AVX/FMA3 ? see no option for it in any version ?

Click to expand...

The problem is what games and productivity application use AVX or SSE SSE2 there is not a list that I know of. I think Crysis3 and handbrake use AVX that I use. You could also turn off SSE SSE2 in prime95 why not and save some more power.

EarthDog said:

if you have a rounding error/instability n p95 stress test, the worker will stop. Tbe rest of the workers keep going, yes.

Perhaps there is a difference in notification delivery (worker stopping or the whole test stopping? However, according to that dev i linked, twice, it's already on.

How do you turn off AVX testing in p95?

Click to expand...

When you enable round off checking it is a extra step in testing it checks with prime numbers and does not stop the worker thread at all from what I see.

You can control whether the program does extra error checking on every
iteration of a primality test. This runs a little bit slower. Windows
user can control these options from the Advanced Menu. Linux users must
add one or both of these lines to prime.txt to control these options:
ErrorCheck=0 or 1

Click to expand...

Stop prime95 before configuring the file.
in local.txt
CpuSupportsAVX=0

What is your source of that quote? That also doesn't reference whether it's the stress test or the program. Clearly you can enable it for the program that we know. Does wherever you sourced that from state if it does anything in the stress tests?

When I have been unstable, threads have dropped and the rest kept going... that's how it works, really. If after x hours it's still using all workers, you are stable.

Any program for video coding x264/x265 will use AVX (i covered this in my 1st post and i personally use Handbrake hence my 2 BIOS profiles) so you should use 28.9 or Realbench. I use a multitude of programs for data-mining, benchs, work related, seti/fold and so far none of them needed AVX. No game as far as i know uses AVX, played Crysis 3 to exhaustion with non-AVX profile and worked perfectly fine. There isn't a list that i know of sadly, but if you start a game and it crashes its usually one of 3 things, overclock, GPU drivers or bad coding

In short there really aren't many programs that actually use AVX, so i specifically repeated what i have been seeing in several forums/threads and thumbed up by mods/devs/programmers - image/video editing, important work like physics or maths, i don't know, government work - 28.9... Everything else 26.6 (or maybe 28.9 with AVX disabled, still have to try it).

Again, f you can run your setup with less voltage, less heat, less wear and tear, smaller electric bill on a daily basis and still be 100% stable for the work/gaming you're doing then why not

EarthDog said:

What is your source of that quote? That also doesn't reference whether it's the stress test or the program. Clearly you can enable it for the program that we know. Does wherever you sourced that from state if it does anything in the stress tests?

When I have been unstable, threads have dropped and the rest kept going... that's how it works, really. If after x hours it's still using all workers, you are stable.

Click to expand...

It says (does extra error checking on every iteration of a primality test) and the Prime.txt file has StressTester when you click on round off checking it is in there and you can turn it off then on. You need to do some testing to see for your self. Have you ever seen a worker thread not stop working and say rounding error in the worker thread?

My resources is the undoc.txt in the prime folder have a read.

I understand what it says. My question though is essentially what does it do that is different? The links I provided earlier are from the developer. He says its already enabled when you stress/torture test. When you actually use Prime the program (not the stress test) it is not enabled. Those instructions from the undoc.txt file are in reference to the running the program not the stress test.

I believe you are saying that it still does something different when you check it off and run the stress test, correct?

This is what it does by default...


This passage is also from the indoc.txt (in the stress testing section) -

The program normally does round-off error checking every 128 iterations.

When you are testing an exponent near the limit of an FFT length, the program
does this error checking every iteration.
NearFFTLimitPct=value
The default value is 0.5 (if the exponent you are testing is within 0.5% of the
maximum exponent that can be tested using the current FFT length, then the
extra error checking is performed). A value of 0.0 will turn off this extra
error checking.

Click to expand...

Yes it is different it is a extra test. When I don't use round off checking I can pass prime for the duration of my testing at a lower voltage. When I do use round off checking on my CPU I have to raise the voltage 0.020 so I don't have non fatal round off checking error, if I can remember correctly it say round off checking error and that is it, also the worker does not stop.

Thanks for the screenshot it made me think I should do one, also that maybe your CPU does not fail like mine. like I was saying before every CPU is different how quickly I forgot..

My i5 6600k PC is down I'm waiting for memory, I will post a screen shot of round off checking error hopefully on Wednesday it will be cool to show.

The program normally does round-off error checking every 128 iterations.

When you are testing an exponent near the limit of an FFT length, the program
does this error checking every iteration.
NearFFTLimitPct=value
The default value is 0.5 (if the exponent you are testing is within 0.5% of the
maximum exponent that can be tested using the current FFT length, then the
extra error checking is performed). A value of 0.0 will turn off this extra
error checking.

Click to expand...

From this quote it looks like some errors will fall through the cracks in favor of performance coming from my quote running extra error checking every iteration of a primality test.

Mixed feelings on 28.9 with CpuSupportsAVX=0, same temps as 26.6 on small FFT but around 10c-15c higher on Blend (hits as high as Small FFT). Otherwise seems to be working well, passed overnight torture. If 28.9 fixes bugs from previous versions (as it should) I might start using this with the tweak.

EDIT: Set local.txt as read-only or it might change and delete CpuSupportsAVX=0

Lots of great information here guys. So far using Prime95 Large FFT or Blend on the newest version just isn't working for me... Im dropping workers before I can get 2 hours of testing passed.

I tried default settings but with manual 1.3 vcore and stressing Large FFTs, one worker went down after less than an hour using my 2x4gb HyperX 2400 MHz sticks too. I can't imagines it's my RAM... unless both sets are bad. Small FFTs don't stress the RAM much tho. Hmmm idk. I did notice my core temps are off more than 10C even tho the ave/max temps are between 50-80C on all tests so far depending on the over lock and voltage. I'm gonna remount when I get a chance to get a small difference in cores.

Prime95 Small FFT works fine though, just cleared 4.5GHz with 1.275V for 5 hours. I'm gonna do more benching tomorrow on Small FFTs. Running 1.250V overnight. So far I got:

4.2GHz - 1.100V for 6 hours with no errors
4.5GHz - 1.275V for 5 hours with no errors

My experience with the new Prime 95 version is limited but it seems to be too extreme for my uses currently

4.5ghz is a very nice overclock for 1.275v, and most don't push it above it. I say you're good right there, enjoy it

wingman99 said:

Yes it is different it is a extra test. When I don't use round off checking I can pass prime for the duration of my testing at a lower voltage. When I do use round off checking on my CPU I have to raise the voltage 0.020 so I don't have non fatal round off checking error, if I can remember correctly it say round off checking error and that is it, also the worker does not stop.

Thanks for the screenshot it made me think I should do one, also that maybe your CPU does not fail like mine. like I was saying before every CPU is different how quickly I forgot..

My i5 6600k PC is down I'm waiting for memory, I will post a screen shot of round off checking error hopefully on Wednesday it will be cool to show.




From this quote it looks like some errors will fall through the cracks in favor of performance coming from my quote running extra error checking every iteration of a primality test.

Click to expand...

Im not sure you are reading and understanding that passage properly.

That passage says it NORMALLY does round off checking (every 128 iterations). It then goes on to say the default value is .5. In other words, by default, its already on. If you change it to 0.0 it will THEN be off. What it DOES say is that when you are running the prime program (again, not the stress test) when you enable rounding errors, it will slow performance. It doesn't mention anything about the stress test there.

Id like to see your screenshot and how it differs from the one I posted. Your words describe, exactly, my picture already...so I am interested to see the difference.

Leegit said:

Lots of great information here guys. So far using Prime95 Large FFT or Blend on the newest version just isn't working for me... Im dropping workers before I can get 2 hours of testing passed.

I tried default settings but with manual 1.3 vcore and stressing Large FFTs, one worker went down after less than an hour using my 2x4gb HyperX 2400 MHz sticks too. I can't imagines it's my RAM... unless both sets are bad. Small FFTs don't stress the RAM much tho. Hmmm idk. I did notice my core temps are off more than 10C even tho the ave/max temps are between 50-80C on all tests so far depending on the over lock and voltage. I'm gonna remount when I get a chance to get a small difference in cores.

Prime95 Small FFT works fine though, just cleared 4.5GHz with 1.275V for 5 hours. I'm gonna do more benching tomorrow on Small FFTs. Running 1.250V overnight. So far I got:

4.2GHz - 1.100V for 6 hours with no errors
4.5GHz - 1.275V for 5 hours with no errors

My experience with the new Prime 95 version is limited but it seems to be too extreme for my uses currently

Click to expand...

@ Leegit - Thanks for being patient with us while we get tot he bottom of this more or less related subject in your thread. When stress testing, I have never enabled this option for round off errors, saw the screenshot I posted that shows round off errors by default. It does not "add floating point" anything. Personally I would run the latest one and just be AVX stable regardless of if you use it or not. I have moved past this program and use AIDA64 stress test and Realbench. That has worked out well for me.

AS far as which to test, only use Small FFT (CPU), or blend (memory and CPU). Large seems to have an inordinately tough time passing. That is decent voltage though for 4.5 GHz!! Nice!

EarthDog said:

Im not sure you are reading and understanding that passage properly.

That passage says it NORMALLY does round off checking (every 128 iterations). It then goes on to say the default value is .5. In other words, by default, its already on. If you change it to 0.0 it will THEN be off. What it DOES say is that when you are running the prime program (again, not the stress test) when you enable rounding errors, it will slow performance. It doesn't mention anything about the stress test there.

Id like to see your screenshot and how it differs from the one I posted. Your words describe, exactly, my picture already...so I am interested to see the difference.

Click to expand...

The program normally does round-off error checking every 128 iterations.

When you are testing an exponent near the limit of an FFT length, the program
does this error checking every iteration.
NearFFTLimitPct=value
The default value is 0.5 (if the exponent you are testing is within 0.5% of the
maximum exponent that can be tested using the current FFT length, then the
extra error checking is performed). A value of 0.0 will turn off this extra
error checking.

Click to expand...

I don't know what normally is or isn't for (every 128 iterations). What if the default is changed from 0.5 to 0.0 that would be 0.0% instead of 0.5%

wingman99 said:

Yes it is different it is a extra test. When I don't use round off checking I can pass prime for the duration of my testing at a lower voltage. When I do use round off checking on my CPU I have to raise the voltage 0.020 so I don't have non fatal round off checking error, if I can remember correctly it say round off checking error and that is it, also the worker does not stop.

Click to expand...

Your screenshot says Fatal Error: rounding was 0.5 expected less than 0.4

Like I was saying in the above post I don't have a screenshot yet my i5 6600k PC is down I'm waiting for memory, I will post a screenshot of round off checking error hopefully on Wednesday it will be cool to show.

Okay so Prime95 Large FFT is tough to pass... got it I'll focus on Small FFT and then try Blend for a bit longer then move on to memory testing.

Been running 4.5GHz @ 1.25V on Small FFT all night currently 9 hours in with zero errors

I was doing some stressing with Prime 95 Small FFT last night at 4.7GHz with 1.325V. I ran it overnight after it cleared the first two hours, woke up this morning (9 hours total run-time) and one worker had stopped after 5 hours from a single error. From what I saw it had around 40-50W power draw throughout the stress but my maximum power draw for the run was 80W. I would guess this spike occurred the moment the worker failed and it had to be resolved by the CPU will still running the other 7 threads.

Has anyone else seen any sort of power draw spike/s like this when pushing a high overclock?