Prime95 Blend, Large FFT fail.

johnrob

Honorable
Nov 22, 2014
100
2
10,695
System:

i5 4690k (currently stock 3.5ghz with 3.9ghz turbo)
CoolerMaster Seidon 240m (aio liquid cooler)
Asus Maximus Hero 6 (z87)
Gskill Sniper Low Voltage (2x4 gig, DDR3 1600, 1.25v)
MSI GTX 970 gaming
EVGA Supernova 850

I tried overclocking before I ever tried the blend test. I OC'd to 4.3ghz running the small FFT test and thought I was okay. I was failing Aida 64 within 30 seconds, and I was getting CTDs in CSGO after that though, So I set everything back to the default settings, then manually set the DRAM frequency and voltage because my BIOS wanted to put it at 1.35v.

I thought it was a DRAM problem so I ran the windows memory diagnostic, extended test 10 passes. It didn't have any errors.

With everything like it is at the moment I can run the small FFT test, AIDA 64 will run, Firestrike runs, pretty much everything runs fine except for the Large FFT test and the blend test.

Both tests have at least 1 core fail instantly. If 2 cores don't fail instantly, the second is short to follow (within 10 seconds) The cores that fail will change, first time it was 1 and 3, then 1 and 4, then 2 and 4, now it's failing on 2 and 3.
 
Solution
As i understand it, the large fft is hell and not supposed to be used for devil's canyon/haswell.

Here's a quote from an intel thermal specialist that posted here on th:

"Section 12 - Thermal Testing @ 100% Workload

Prime95 Small FFT's is the standard for CPU thermal testing, because it's a steady-state 100% workload. This is the test that Real Temp uses to test sensors. The link above is to version 26.6, which is well suited to all Core 2 and Core i variants.

Core i 2nd, 3rd and 4th Generation CPU's have AVX (Advanced Vector Extension) instruction sets. Recent versions of Prime95 run AVX code on the Floating Point Unit (FPU) math coprocessor, which produces unrealistically high temperatures. The FPU test in the software utility...

johnrob

Honorable
Nov 22, 2014
100
2
10,695


I have not manually set the timings, they all looked correct though? (9-9-9-24) If there's more to it than that then I'll have to read up on memory settings for I do not know a lot about it.

@1.35v AIDA 64 was failing within 30 seconds, and prime was giving the same results. I experienced one BSOD on starting the large FFT test however.
 
As i understand it, the large fft is hell and not supposed to be used for devil's canyon/haswell.

Here's a quote from an intel thermal specialist that posted here on th:

"Section 12 - Thermal Testing @ 100% Workload

Prime95 Small FFT's is the standard for CPU thermal testing, because it's a steady-state 100% workload. This is the test that Real Temp uses to test sensors. The link above is to version 26.6, which is well suited to all Core 2 and Core i variants.

Core i 2nd, 3rd and 4th Generation CPU's have AVX (Advanced Vector Extension) instruction sets. Recent versions of Prime95 run AVX code on the Floating Point Unit (FPU) math coprocessor, which produces unrealistically high temperatures. The FPU test in the software utility AIDA64 shows the same results.

It's not necessary to run AVX code for thermal testing. Prime95 v26.6 produces temperatures on 3rd and 4th Generation processors more consistent with 2nd Generation, which also have AVX instructions, but do not suffer from thermal extremes due to having a soldered Integrated Heat Spreader and a 35% larger Die.

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."
 
Solution

CompuTronix

Intel Master
Moderator


You left out the operative words; for CPU Thermal Testing.

"Prime95 Small FFT's is the standard for CPU thermal testing, because it's a steady-state 100% workload."

A steady-state workload is necessary to analyze thermal performance at 100%. This means that the temperatures are steady ... not fluctuating.

Shown above from left to right: Small FFT's, Blend, Linpack and Intel Burn Test.

Shown above from left to right: Small FFT's, Intel Extreme Tuning Utility CPU Test, and AIDA64 CPU Test.

The "Charts" in SpeedFan span 13 minutes, and show how each test creates different thermal signatures. Intel Extreme Tuning Utility is also a cyclic workload. Although AIDA64's CPU test is steady-state, the workload is insufficient, and it's not available as freeware.


You can run whatever you like for Stability Testing, but Small FFT's is appropriate for Thermal Testing.

CT :sol:
 
I didn't leave out for thermal testing, it was in the section I quoted. The op said "I OC'd to 4.3ghz running the small FFT test and thought I was okay" - "With everything like it is at the moment I can run the small FFT test, AIDA 64 will run, Firestrike runs, pretty much everything runs fine except for the Large FFT test and the blend test." Thermals are relevant since too much heat can cause shutdowns. Run a test known to cause excessive heat and it causes problems, it's somewhat to be expected. I was just sharing further info as to why this was potentially happening and it's due to the design of the chip becoming super stressed with the on die math coprocessor adding significant heat to what should be a standard stress test for older chips.
 
Some additional info regarding haswell and prime95..

http://www.overclockers.com/forums/showthread.php/751197-Prime-95-v28-5-Haswell-CPU-s-and-you-is-it-bad-for-the-CPU-s-health

There also seem to be issues with running prime95 using 'adaptive' (most oc'ers suggest turning adaptive off) and failing to have the voltages set to manual. The haswells are a bit worse for it than previous chips and have a habit of cranking up the voltages under heavy stress loads when the cpu ramps up, in turn causing wicked heat when voltages are left on 'auto' and it's left to it's own devices.
 
I wasn't implying that temp is the only thing to test or check for but the two are definitely related. If you're testing for stability (voltage handling under load) and in doing so cause temps to skyrocket beyond the thermal design, that's unstable. The two have to coexist nicely. Heat will cause shutdowns just as much as voltage issues.
 

CompuTronix

Intel Master
Moderator
synphul,

Yes, that's why this is in the Temp Guide:

Section 8 - Overclocking and Vcore

" ... Overclocked processors can reach up to 150% of their Thermal Design Power (TDP) when using manual Core voltage (Vcore) settings, so high-end air or liquid cooling is critical. Every processor is unique in it's overclocking potential, voltage tolerance and thermal behavior.

Regardless, excessive Vcore and temperatures will result in accelerated "Electromigration" - https://www.google.com/?gws_rd=ssl#q=Electromigration - which prematurely erodes the traces and junctions within the processor's layers and nano-circuits. This will eventually result in blue-screen crashes, which will become increasingly frequent over time.

CPU's become more susceptible to Electromigration with each Die-shrink, so 22 Nanometer architecture is less tolerant of over-volting. Nevertheless, Vcore settings should not exceed the following:

-> Core 2

1st. Generation 65 Nanometer ... 1.50 Vcore
2nd Generation 45 Nanometer ... 1.40 Vcore

-> Core i

1st. Generation 45 Nanometer ... 1.40 Vcore
2nd Generation 32 Nanometer ... 1.35 Vcore
3rd Generation 22 Nanometer ... 1.30 Vcore
4th Generation 22 Nanometer ... 1.30 Vcore ... "


I realize that the topics of processor temperatures and overclocking overlap considerably, and there is much material and many details that I have chosen to leave out of the Temp Guide. A few years ago, the previous version was 16 pages which included an entire Section devoted to calibrations. Some users suggested that the Guide was too lengthy and too detailed, so I chose to make this new version more concise.

Is there anything specific that you'd like to see added to the Guide?

CT :sol: