Haswell-E impossible to get it prime95 stable with a reasonable OC?

AaA91 · 19 Apr 2016 at 19:14

Hi Guys!

Those are my specs :-
i7 5820K @ 125x35 Core, 125x33 Cache
Asus X99-Pro/ USB 3.1
16 GB Corsair Veng. LPX DDR4-3000

I am trying to be prime95 28.7 stable with a reasonable OC and have a very high difficulty of achieving this.

First I tried 125x35 @ 1.240V and it crashed in smallFFTs in about 35mins.
Increased Vcore to 1.255 and it crashed just shy of 2 hours. I increased it to 1.270V and was able to pass 4.5 hours then I moved to largeFFts (1344k), then It froze in less than 35 mins. I thought it might be the RAM so I overvolted it to 1.4 instead of 1.35 and froze after about 1.5 hours, I tried loosen the timings a tad and it crashed after 1 hour. I am now thinking it is the cache that is responsible for the freezes so I increased it to 1.25V instead of 1.21V. And I will see if it passes without issues.
(Edit: bam actually it froze again at the time of writing this thread, lol, so probably it is not the cache but something else)

I am really tired of this, it is really giving me headaches. My system otherwise passed Realbench for 8 hours, Aida64, OCCT (Small Data and Large Data Sets). So I tried prime95 as a final test and it failed miserably at it.
What concerns me is any of those programs are not using any of the new instructions, and when programs showed up using them I will have instability issues just like prime95.

reef · 19 Apr 2016 at 19:24

Use v26.6 for Haswell as 28.7 uses AVX and heats up more and requires more voltage to be stable.

On my 4770K @ 4.5 temps hit 100C within minutes on 28.7 as the voltage required and heat created is much higher than any other application I run.

tyler_jrb · 19 Apr 2016 at 19:35

Use realbench. Prime95 is just too demanding and nothing realistic at all especially for haswell-e 6/8 cores.

If you must record a video and compress it on handbrake.

Up your TDP in the bios aswell to ensure its able to use as much power as it needs on full load. Should be something like "long duration power limit" under cpu power management.

zipp0r · 19 Apr 2016 at 22:17

try to overclock each thing seperately and test it.

is 3000mhz ram officially supported? if not, you're running the memory controller out of spec and vccio will probably need increasing from 1.05 to 1.10 to pass large ftt's.

Borealis · 19 Apr 2016 at 22:31

Anything over 2133MHz (the official standard) is an overclock, but if you buy a 3000MHz kit, that means that it has been tested at that speed (I could be wrong). Of course, every memory controller is different and yours might need a bump in voltage to get it stable.

AaA91 · 19 Apr 2016 at 23:53

Thanks for the replays guys. I managed to stabilize it, all this acting up is because the CPU is lacking some VCCIN

joeyjojo · 20 Apr 2016 at 08:00

Well done OP. What have you set VCCIN to out of interest? Still good with the memory at full speed?

reef said:
Use v26.6 for Haswell as 28.7 uses AVX and heats up more and requires more voltage to be stable.

tyler_jrb said:
Use realbench. Prime95 is just too demanding and nothing realistic at all especially for haswell-e 6/8 cores.

If you must record a video and compress it on handbrake.

This is bad advice. What if you run AVX maths solvers all day? Or prime95 looking for prime numbers? They're just as "realistic" a load as any other (what does this even mean??)

And replacing it with video encoding is silly, it's got no error detection (other than crashing).

Borealis · 20 Apr 2016 at 08:32

Interesting results. What VCCIN are you using as it might help me work out why my 5820K isn't stable above 125x33 as well?

tyler_jrb · 20 Apr 2016 at 09:01

joeyjojo said:
Well done OP. What have you set VCCIN to out of interest? Still good with the memory at full speed?

This is bad advice. What if you run AVX maths solvers all day? Or prime95 looking for prime numbers? They're just as "realistic" a load as any other (what does this even mean??)

And replacing it with video encoding is silly, it's got no error detection (other than crashing).

I've ran aida64, Intel burn, real bench. With my old cooler. In the low 70s. Prime 95. Over 90c in seconds. It is not realistic for a standard usage scenario. And will just result in overloading the CPU for no reason. Depending on what version of prime, I've known CPUs die because of it.

Is OP using his PC for AVX? I doubt it. And even then it's not as harsh as prime is.

I've ran hours on end with prime, aida64 then compressed a few videos only to have it crash due to instability on handbrake.

joeyjojo · 20 Apr 2016 at 09:19

tyler_jrb said:
I've ran aida64, Intel burn, real bench. With my old cooler. In the low 70s. Prime 95. Over 90c in seconds. It is not realistic for a standard usage scenario. And will just result in overloading the CPU for no reason. Depending on what version of prime, I've known CPUs die because of it.

Is OP using his PC for AVX? I doubt it. And even then it's not as harsh as prime is.

I've ran hours on end with prime, aida64 then compressed a few videos only to have it crash due to instability on handbrake.

I think you're a bit confused.

I don't know about aida or realbench, but recent Intel Burn Test (or linx) uses AVX. If used properly it should get hotter than anything else, even Prime95 small FFTs. It's very sensitive to the problem size so you need to experiment with that.

As for "Is OP using his PC for AVX? I doubt it" - plenty of applications are compiled with AVX(2) extensions. Handbrake (specifically the x264 encoder) is one. It's highly optimised and will get things hot. The problem with using it as a stability test is it won't flag any errors (other than catastrophic ones), unlike Prime95 or IBT/linx.

If you think x264 is a "realistic" load then so are IBT and Prime95.

zipp0r · 20 Apr 2016 at 11:21

AAA91, why only 1344k fft's ? i run either blend and it will cycle through all ftts in about 24hrs or i use the largest 4096 ftt

zipp0r · 20 Apr 2016 at 11:29

tyler_jrb said:
I've ran hours on end with prime, aida64 then compressed a few videos only to have it crash due to instability on handbrake.

didnt run it for long enough :confused:

24hrs small ftt, 24hrs blend/4096k ?

joeyjojo said:
I think you're a bit confused.

I don't know about aida or realbench, but recent Intel Burn Test (or linx) uses AVX. If used properly it should get hotter than anything else, even Prime95 small FFTs.

for me, ibt will hit the same temps that p95 small ftt with avx will hit. occt linpack avx will get the cpu even hotter tho, not sure why, but easily another +5c :eek:

joeyjojo · 20 Apr 2016 at 11:37

zipp0r said:
for me, ibt will hit the same temps that p95 small ftt with avx will hit. occt linpack avx will get the cpu even hotter tho, not sure why, but easily another +5c

Interesting, maybe different problem sizes? In my testing the time taken per loop, the GFLOPs, and the heat all increase with problem size up to a point. After that it just takes longer without being any more intensive.

It looks like OCCT uses 90% of memory by default, which is a lot more than the medium setting on IBT. If you set them like for like they should be the same. As long as you've got AVX on both obvs.

AaA91 · 20 Apr 2016 at 15:39

joeyjojo said:
Well done OP. What have you set VCCIN to out of interest? Still good with the memory at full speed?

.

Thanks. I tested it for 4 hours and it was stable during the test with 95% of the memory loaded, the memory speed was at 3000 MHz.

Borealis said:
Interesting results. What VCCIN are you using as it might help me work out why my 5820K isn't stable above 125x33 as well?

I set VCCIN to 1.930 in BIOS LLC8 and it can reach 1.952 sometimes during stress tests.

zipp0r said:
AAA91, why only 1344k fft's ? i run either blend and it will cycle through all ftts in about 24hrs or i use the largest 4096 ftt

From what I've heard 1344K/1792K FFts are the hardest for the Haswells. Even harder than 4096K.

Borealis · 20 Apr 2016 at 15:47

Thanks for the info. I'll check and see what my VCCIN is tonight. I take it LLC8 is the load-line calibration level? I'm using a Gigabyte board, so my LLC is described as normal, high, extreme etc.

AaA91 · 20 Apr 2016 at 15:53

Borealis said:
Thanks for the info. I'll check and see what my VCCIN is tonight. I take it LLC8 is the load-line calibration level? I'm using a Gigabyte board, so my LLC is described as normal, high, extreme etc.

Yeah LLC = Load-Line Calibration. According to BIOS description it runs at 100% level (There is one level higher that is Level 9 and it is @ 125%), so yours at high is probably equivalent to LLC 8 on my Asus mobo.

Good Luck

Borealis · 20 Apr 2016 at 16:07

Thanks.

Yeah, I'll have a look. TBH it's only another 250MHz, but I know so many people with 5820Ks above 4.1GHz that I think I must be doing something wrong. Unless I've got a poor chip of course.

~>Dg<~ · 20 Apr 2016 at 16:23

its the luck of the dip isnt it.

some struggle older ones at even 4ghz yet newer versions can do 4.5 easy.as long as you in the ball park and not benchmarking 24/7 what does 100-200mhz matter.

mmj_uk · 20 Apr 2016 at 16:34

Have you tried overclocking without the 1.25 strap? that might be causing things to become unstable?

WhiteHawk77 · 25 Apr 2016 at 02:57

As far as I was aware Prime95 is not recommended on Haswell and that's what I was told here when I first OCed my 5930K which failed and which is now dead 7 months later. I also saw Linus mention in a video that it's known as a power virus and can actually damage CPU's. 89c ouch, what you cooling it with?