BSOD's with Prime95 ver. 27.4

[flyfisher]

System in sig, was always stable with linx, Prime95 26.6, HCI memtest. Now with the Prime 27.4 that uses AVX I eventually get a BSOD with large FFTs, either 101 or 124 which indicates not enough vcore. I increased Vcore a notch and it still BSODs. Is this just that more stressful on the CPU?

The other thing that I noticed is that load voltage is actually higher running 27.4 than 26.6. I know software monitoring isn't always accurate but why would there be a difference?

[Chike]

Re: BSOD's with Prime95 ver. 27.4

Using AVX means if double floating poiny ops so of course there is more "stress" on the CPU and memory.
Might be Vccio(VTT) or Vccsa as well.

[Darkje]

Re: BSOD's with Prime95 ver. 27.4

I remember a remark that in prime 27 the load shifted from cpu to imc, or something to that effect, I'll try and find the reference...

edit:
yes found it, from the 27.3 release notes:
"... This version supports 64-bit optimized AVX FFTs. 32-bit AVX FFTs are also a little bit faster. ...
The good/bad news is these FFTs are so fast that they are limited by memory bandwidth -- standard Sandy Bridge CPUs will experience a slow down when running all 4 cores. I'd like to hear from Sandy Bridge-E users to see if they also suffer slow downs when all 4 cores are running."

http://www.mersenneforum.org/showthread.php?t=16535

so perhaps in your case the imc is hammered so hard that it needs more juice...

[flyfisher]

Re: BSOD's with Prime95 ver. 27.4

Thank you for the information. I am not going to worry about it then, the machine is perfectly stable for everything that I do. Although I must admit I don't understand, if Sandy Bridge has such a high memory bandwidth, more than it can use, hence why DDR3 1600 is the so-called sweet spot, how is does the CPU experience a slow-down due to limited memory bandwidth?

[Darkje]

Re: BSOD's with Prime95 ver. 27.4

While I don't know the specifics I guess the avx instructions make the calculations so fast that the memory does get swamped by the many requests. Your observation of higher vcore (or less droop) seems to be in line with that, [/guess mode] so cpu not taking as much power as before, not calculating but slightly delaying? [/guess mode off]

[flyfisher]

Re: BSOD's with Prime95 ver. 27.4

Thanks, I believe you are correct in your assessment. That would explain the higher Vcore (or less Vdroop) as you said. I wonder how a Sandy Bridge-E would handle it with quad channel.

[wevsspot63]

Re: BSOD's with Prime95 ver. 27.4

I've been playing with my GA-Z68XP-UD4 (revision 1.3) motherboard for a few weeks now. Like you, I couldn't pass 12 hours of the latest version of Prime95. Tweaked and tweaked, tested and retested and frankly was about ready to toss the board in the trash can because I couldn't seem to get it to pass Prime95 for the life of me.

I found three key things that finally got me toaster stable;

1. Increased VCCIO voltage to 1.10v (16Gb of G.Skill DDR3 1866Mhz)
2. Decreased CPU PLL voltage to 1.76v
3. Manually set VCCSA voltage to .925v

Now running happily at 4.5Ghz with all power saving features enabled and using DVID offset voltage control (also enabled Multi Steps LLC using TouchBios - this option is disabled in the Award BIOS if you enable normal voltage control). For me LLC level 4 with a DVID offset of +.04v was the sweet spot. This yields a CPUz and TouchBios reading of 1.32v under OCCT and Prime95 loads.

Before anyone gets onto me about changing the VCCSA voltage, in the BIOS version I'm running (F6) VCCSA is auto set at .920v. However, according to the Intel specs. the nominal value is .925v, accordingly I manually set mine to the Intel recommended nominal value.

I can probably drop my offset voltage by .01 or even .02 to get 1.30v to 1.308v cpu vcore under load and still remain stable. Just haven't had the time yet to find the lowest voltage needed to run at 4.5Ghz Prime95 stable. That's next up on my to do list, as well as see if I can hit 4.6Ghz stable for 24/7 use.

[flyfisher]

Re: BSOD's with Prime95 ver. 27.4

Thanks for the info. My board is a rev. 1.0, but must use the same F6 BIOS. I also am using DVID, but no LLC. I have VCCIO at 1.075, and VCCSA at 0.925. I am playing around with 2x4GB 30nm Samsung low profile low voltage memory now, will see if I can get it stable with ver. 27.4.

[wevsspot63]

Re: BSOD's with Prime95 ver. 27.4

Good luck Flyfisher. Hopefully the new RAM will play pretty.

For me the ability to utilize some level of LLC was really key. Even though I've got a great (and plenty powerful) power supply - Corsair HX1000, I was still seeing major vdrop and vdroop when putting the system under heavy stress. For example, without LLC with BIOS voltage set at 1.36 my software reported load voltage drooped to 1.288 - 1.300v. And it was a weird droop too. Using OCCT logging to watch my vcore, it would first droop down to 1.32v and then after approx. 30 seconds drop down again to 1.288 - 1.30v and sit pretty steady there except when the worker threads were changing.

At least in my situation I can confidently say that using a small amount of LLC (lvl 4), decreasing CPU PLL voltage, increasing VCCIO and VCCSA just a smidge resulted in a perfectly stable system. I can't express enough how frustrated I was with this board the first week and a half. And when it failed Prime95 it wasn't just a worker thread stopping either, it was a full blown BSOD every single time. Error code was the unfortunately common x124.

[flyfisher]

Re: BSOD's with Prime95 ver. 27.4

Yay! I got it stable with large FFTs for 8 hours at 4.3 GHz, memory 1866 MHz @9-9-9-27-1T, VDIMM 1.375, VCCIO 1.075, VCCSA 0.930, PLL 1.80, DVID -0.010, no LLC. Vcore shows 1.32 in BIOS in the main MIT screen. In Windows CPU-Z shows 1.296 Vcore under load with ver. 27.4, 1.284 with ver. 26.6. Now to try 4.4 GHz. Thanks for the help.

[Lsdmeasap]

Re: BSOD's with Prime95 ver. 27.4

Great to see some progress and success here! I haven't even tried this version yet, so I guess I'll have to check it out sometime soon.

I've been a 5-10 LinX and HCI memtest guy for a long time, haven't seen any need for Prime95 testing anymore lately but I guess I'll check this out now that it's been properly updated for newer CPU's

[Darkje]

Re: BSOD's with Prime95 ver. 27.4

... In Windows CPU-Z shows 1.296 Vcore under load with ver. 27.4, 1.284 with ver. 26.6.

yeah, 27.4 is less droop on vcore compared to 26 and linx... this shift in behaviour is making it an other test than it used to be, more imc/memory critical (and perhaps less cpu?)...

[Lsdmeasap]

Re: BSOD's with Prime95 ver. 27.4

Nice, I'll try it out sometime again! I just don't have time for 8+ hours of testing, so I doubt I'll ever use it for any frequent stability testing purposes

[wevsspot63]

Re: BSOD's with Prime95 ver. 27.4

Glad to see that you're making progress. Congratulations!

[wevsspot63]

Re: BSOD's with Prime95 ver. 27.4

The extended runs of Prime95 were the ones really getting to me. I could pass LinX, OCCT, HCI perfectly but get 3-5 hours into a Prime95 run and BSOD. Not that I'd ever have my ram, cpu and imc under full load for 5 hours, but still comforting to know that it's toaster stable when passing all of the stress tests including a nice long 8 or more hour run of Prime95.