4690K and Prime95 26.6 vs 28.5

Stuka87

Diamond Member
Dec 10, 2010
6,240
2,559
136
Hey All,

So building a new system and was doing some CPU testing last night to test temps and over clocks. And my first results with Prime95 28.5 worried me (Small FFTs), but then after some reading people say not to use this version on Haswells.

So I downloaded the older 26.6 and reran my test and go way lower temps. But I am really questioning which one to believe.

4690K @4.3GHz @1.15v (have not tried higher clocks yet)

Idle temps: 26-28C in 23C ambient temp

Prime95 26.6 (SmallFFT): After 20 mins it hit 58C

Prime95 28.5 (SmallFFT): After 20 mins it hit 80C

So the big question is... what version and which temp should I trust? Which such a huge difference between them I am not sure I want to trust either? Is there another tool that I should use instead of Prime95 (Which I have used for years, but this is my first Haswell)?

Thanks!
 

waffleironhead

Diamond Member
Aug 10, 2005
7,122
622
136
After 26.6, prime95 introduced AVX optimizations. that's the difference. If you aren't running any avx code, you might never reach those operating temps.
 

flexy

Diamond Member
Sep 28, 2001
8,464
155
106
AVX2 instructions add 0.15 VCore, whether you want it or not..so those tests get CRAZY hot. I would simply avoid them since I don't see them as anywhere remotely realistic.
 

Stuka87

Diamond Member
Dec 10, 2010
6,240
2,559
136
Ah, that explains it then. If it increases the vcore, I was really at 1.3v.

So then the 58C I was seeing with 26.6 is far more realistic for real world use. Unless games or something start using AVX code?
 

flexy

Diamond Member
Sep 28, 2001
8,464
155
106
Even then you will never see those temps "in real life"...I think a far more realistic test is maybe X264 benchmark or ASUS Real Bench.
 

ShintaiDK

Lifer
Apr 22, 2012
20,378
146
106
AVX2 instructions add 0.15 VCore, whether you want it or not..so those tests get CRAZY hot. I would simply avoid them since I don't see them as anywhere remotely realistic.

100mv, not 150mv. And its all 256bit AVX instructions. Cache speed is also doubled to ~1TB/sec.

Performance also gets a ~80% boost.
 

DrMrLordX

Lifer
Apr 27, 2000
23,223
13,303
136
AVX2 instructions add 0.15 VCore, whether you want it or not..so those tests get CRAZY hot. I would simply avoid them since I don't see them as anywhere remotely realistic.

That's a sad statement, unfortunately. There's probably a fair amount of software out there that could really take advantage of AVX2 that isn't doing so yet.

Maybe someday it will, just like SSE3/4 support is pretty widespread nowadays.
 

ShintaiDK

Lifer
Apr 22, 2012
20,378
146
106
It doesnt get hot due to AVX usage, it gets hot due to almost 100% AVX usage with maximum throughput. But again, in this case a 4670 is around 80% faster than a 3570.