The Reality of Haswell Overclocking - Results Poll

[wabbitslayer]

4.4 @ 1.25 adaptive voltage, never seen temps above 65 while doing gaming, working, doing "real world" stuff. Never gone above 75 doing any sort of 24 hour stress-testing that doesn't involve avx; I set OCCT to stop testing at 90, it has done it every time within a few minutes using avx.

I know the leet overclockers and purists say this isn't a 24/7 "stable" overclock, but I have run mine 24/7 since July and not one bsod/crash. Good enough for me...a calculation or two may have been made in error, but I'm not splitting atoms or performing brain surgery.

Everyone loves to blame the IHS and say Intel screwed that up, and I guess the results of people who delid get bears that out, but maybe it's the avx instruction set itself that Intel messed up.

 

[Kougar]

I could do 3.9 on stock voltage with thermals in the high 50's on a dual 120 rad while running IBT...something seems off.

Again, it originally ran fine at 4Ghz using stock settings. Temps were somewhere below 60 at stock settings, I didn't pay attention to those until I began adding voltage later on.

A few months at 4.2Ghz and the chip began degrading, meaning I had to either raise the voltages or lower the clocks to keep it passing linpack tests. I first raised the voltages, then finally was forced to back the core/uncore both to 4Ghz at the same voltages I'd been using at 4.2Ghz and that worked for awhile. Eventually even that also began to fail with increasing frequency.

Processors degrading is nothing new and some steady 24/7 F@H loads at 1.26vCore and ~1.26v Ring voltage was enough to do it for my 4770K chip. I followed guides and even copied three other GB Z87 user settings just to be sure before I threw in the towel on that launch day chip.

 

[Termie]

There's another thing we haven't covered in this thread. Overclocking results are significantly affected by cache speed and RAM speed. Unless stated otherwise, we should probably assume overclocks are at default cache clocks and DDR3-1600.

 

[24601]

The people saying they got stable above 4.3 ghz without delidding are mostly lieing to themselves and others.

The general seething hatred against obvious stability tests is proof enough.

The same people hate running furmark with an unlimited bios on their GPUs as it shows their limits there as well.

 

[abbcccus]

I've overclocked everything I've ever owned that could be overclocked, at the very least to see what it could do. I've always settled, though, on a modest overclock that kept the voltage close to stock since my office is small and it gets hot very easily (I'm looking at you Q6600 - only quad I owned until Lynnfield as it raised, no kidding, my office temperature five freaking degrees at 3.0 GHz!). My 4770k is the first time I've ever pushed it as far as I could. I'm limited by cooling, I think, but a Zalman CNPS9900 Max isn't exactly chopped liver, so maybe I could get another .1 or .2? Anyhow, 4.3 GHz at 1.246 volts. Temps start creeping awfully close to 100 at 4.5 and even 4.4, but Prime95 crashes at those frequencies. IBT is fine, oddly enough, at the higher frequencies. At least Prime95 had the decency to crash after only an hour or so.

 

[SlowSpyder]

Nevermind.

 

[solidsnake1607]

I have a Core i5 4670k (Costa Rica) Oced to 4.4 at 1.15v hitting 68º in the hottest core. Quite happy with my chip. I think it can do 4.6GHz at 1.2 but i'll wait for the Corsair H105 to give it a shot! Anything above 1.3 is too high for a 22nm chip so it will overheat. If you are an overclocker and want to go above 4.8 you'll have to delid the chip or sell yours and look for a good one which can operate at very high frequencies with low voltage. I owned a 4770K delided which hits 4.9 at 1.34v and i regret selling it, i sold it for 400usd to a overclocker guy... Im not a big overclocker anyways so i dont feel too bad about it.

 

[Grooveriding]

Haswell is an OC dud. Roasts like fire and generally sees poor results. I think HW made sense for non overclockers. But if you had a nice SB or IB, after overclocking you'd never see the difference. My 4670K does 4.3 stable with almost 1.8V, after that you need massive voltage and it roasts.

My worry is Intel's successor to Haswell may wind up just being power use improvements with no performance gains to speak of, maybe 5%, and overclocks like an even worse pig than HW is.

 

[rtsurfer]

The people saying they got stable above 4.3 ghz without delidding are mostly lieing to themselves and others.

The general seething hatred against obvious stability tests is proof enough.

The same people hate running furmark with an unlimited bios on their GPUs as it shows their limits there as well.

Not everyone believes in make shift stability.

4.6 Ghz @1.315V. No delid, H100i.

I dont think it gets tougher than AVX2 Linpack.

 

[ShintaiDK]

Its the same story over again. Back then some SB owners trashtalked IB. Now some IB owners trashtalk HW. All to feel better of having last gen tech I guess.

 

[24601]

Not everyone believes in make shift stability.

4.6 Ghz @1.315V. No delid, H100i.

I dont think it gets tougher than AVX2 Linpack.

Balla got 220+ gflops @ 4.6ghz.

Whatever you're doing, its not 4.6ghz.

http://forums.anandtech.com/showpost.php?p=35159373&postcount=45

I'm not going to rehash that entire thread again.

Took too much of my effort just to do it the first 10 times.

I like how you are using subzero ambients though, truly sportsmanlike.

 

[rtsurfer]

Balla got 220+ gflops @ 4.6ghz.

Whatever you're doing, its not 4.6ghz.

It is the HT hit.
If I turn it off the score should go up.
I even got 170Gflops+ when i turned my RAM up at it XMP speed of 2400.
I didnt take a screenshot at that point as I didnt think I would need it.

Could do a re-run sometime next week if you want to see it.
I've got 4.6Ghz & you can't deny it.


Edit: I know the ambients issue was going to be raised.
I put my PC outside in my balcony where according to the weather app, the ambient should have been -7ºC.
The reason is because I couldnt get past 4.5 Ghz inside my house & I dont have a custom loop.

Regardless, the CPU does not go above 70 ºC in my house when I am gaming.

Edit2: Linpack is atleast 10 degrees warmer for me than Prime95 small FFT. I prefer to use Linpack as my stability test as it works quickly. I dont want to run prime95 for 24 hours only to see it crash after 18hours. That is just stupid. Linpack does the same thing, just quicker. We all know how stability test are way hotter than real world applications, but I still wanted to pass Linpack so I took it outside, not in a temperature i could survive in, but also running the kind of load I would never have used.

 

[24601]

It is the HT hit.
If I turn it off the score should go up.
I even got 170Gflops+ when i turned my RAM up at it XMP speed of 2400.
I didnt take a screenshot at that point as I didnt think I would need it.

Could do a re-run sometime next week if you want to see it.

Edit: I know the ambients issue was going to be raised.
I put my PC outside in my balcony where according to the weather app, the ambient should have been -7ºC.

The reason is because I couldnt get past 4.5 Ghz inside my house & I dont have a custom loop.

Regardless, the CPU does not go above 70 ºC in my house when I am gaming.
I've got 4.6Ghz & you can't deny it.

The LN2 people don't delid either.
They totally count as well obviously.

 

[rtsurfer]

The LN2 people don't delid either.
They totally count as well obviously.

I can run 4.5 @ 1.275 V inside my house. The thing is I didnt have a custom loop. My CPU is limited by its cooling, not its capabilites, I took it outside so I could somehow emulate the kind of Cooling I would have had with a Custom loop.

Believe me, my CPU boots at 4.8 @ 1.2 V and runs good outside games & stability tests. I could boast around telling everyone that, but I too am Interested in proper stability hence the Linx tests.

 

[OCGuy]

Its the same story over again. Back then some SB owners trashtalked IB. Now some IB owners trashtalk HW. All to feel better of having last gen tech I guess.

Bingo

 

[Tristor]

The people saying they got stable above 4.3 ghz without delidding are mostly lieing to themselves and others.

Even with a delid and AIO water, there's a huge voltage wall at 4.4, and another at 4.6 for me on what by all accounts is a pretty cherry chip. To be fair though, if you're getting to at least 4.4, you've got nothing to complain about. That's a 1GHz overclock over stock. Maybe you're not setting world records on the weekend, but really who ever has? 4.4 on HW is also plenty fast enough to make you no longer CPU limited in any games.

Balla got 220+ gflops @ 4.6ghz.

Whatever you're doing, its not 4.6ghz.

http://forums.anandtech.com/showpost.php?p=35159373&postcount=45

I'm not going to rehash that entire thread again.

Took too much of my effort just to do it the first 10 times.

I like how you are using subzero ambients though, truly sportsmanlike.

Nice link, I decided to see what it'd look like for me. First though, I wanted to establish a baseline at 4.4. I'll do 4.6 later, I have some other stuff to do today so won't be able to retest again until tonight.

Intel(R) Optimized LINPACK Benchmark data

Current date/time: Sat Feb 08 11:22:17 2014

CPU frequency:    4.398 GHz
Number of CPUs: 1
Number of cores: 4
Number of threads: 4

Parameters are set to:

Number of tests: 12
Number of equations to solve (problem size) : 1000  2000  3000  4000  5000  10000 15000 20000 25000 30000 35000 40000
Leading dimension of array                  : 1000  2000  3000  4000  5000  10000 15000 20000 25000 30000 35000 40000
Number of trials to run                     : 4     4     4     4     4     2     2     2     2     1     1     1    
Data alignment value (in Kbytes)            : 4     4     4     4     4     4     4     4     4     4     4     4    

Maximum memory requested that can be used=4210869504, at the size=40000

=================== Timing linear equation system solver ===================

Size   LDA    Align. Time(s)    GFlops   Residual     Residual(norm) Check
1000   1000   4      0.006      115.6602 1.083189e-012 3.693953e-002   pass
1000   1000   4      0.006      115.3749 1.083189e-012 3.693953e-002   pass
1000   1000   4      0.006      113.5161 1.083189e-012 3.693953e-002   pass
1000   1000   4      0.006      121.3320 1.083189e-012 3.693953e-002   pass
2000   2000   4      0.043      123.8458 4.220901e-012 3.671667e-002   pass
2000   2000   4      0.043      124.9015 4.220901e-012 3.671667e-002   pass
2000   2000   4      0.045      119.4201 4.220901e-012 3.671667e-002   pass
2000   2000   4      0.041      130.8369 4.220901e-012 3.671667e-002   pass
3000   3000   4      0.129      139.9826 1.016483e-011 3.914231e-002   pass
3000   3000   4      0.142      126.7914 1.016483e-011 3.914231e-002   pass
3000   3000   4      0.127      142.3348 1.016483e-011 3.914231e-002   pass
3000   3000   4      0.126      142.6667 1.016483e-011 3.914231e-002   pass
4000   4000   4      0.294      145.2426 1.906425e-011 4.155234e-002   pass
4000   4000   4      0.270      158.2924 1.906425e-011 4.155234e-002   pass
4000   4000   4      0.288      148.4005 1.906425e-011 4.155234e-002   pass
4000   4000   4      0.265      161.3818 1.906425e-011 4.155234e-002   pass
5000   5000   4      0.539      154.8257 2.299338e-011 3.206242e-002   pass
5000   5000   4      0.497      167.7242 2.299338e-011 3.206242e-002   pass
5000   5000   4      0.544      153.1397 2.299338e-011 3.206242e-002   pass
5000   5000   4      0.503      165.8536 2.299338e-011 3.206242e-002   pass
10000  10000  4      3.448      193.3957 9.420734e-011 3.321846e-002   pass
10000  10000  4      3.464      192.5038 9.420734e-011 3.321846e-002   pass
15000  15000  4      10.897     206.5210 2.137378e-010 3.366406e-002   pass
15000  15000  4      11.145     201.9278 2.137378e-010 3.366406e-002   pass
20000  20000  4      24.719     215.7877 3.723286e-010 3.295924e-002   pass
20000  20000  4      24.415     218.4748 3.723286e-010 3.295924e-002   pass
25000  25000  4      47.560     219.0496 5.678848e-010 3.229357e-002   pass
25000  25000  4      47.617     218.7873 5.678848e-010 3.229357e-002   pass
30000  30000  4      80.943     222.4017 7.893267e-010 3.111534e-002   pass
35000  35000  4      130.546    218.9711 1.216992e-009 3.532743e-002   pass
40000  40000  4      199.288    214.1110 1.471200e-009 3.272000e-002   pass

Performance Summary (GFlops)

Size   LDA    Align.  Average  Maximal
1000   1000   4       116.4708 121.3320
2000   2000   4       124.7511 130.8369
3000   3000   4       137.9439 142.6667
4000   4000   4       153.3293 161.3818
5000   5000   4       160.3858 167.7242
10000  10000  4       192.9497 193.3957
15000  15000  4       204.2244 206.5210
20000  20000  4       217.1312 218.4748
25000  25000  4       218.9184 219.0496
30000  30000  4       222.4017 222.4017
35000  35000  4       218.9711 218.9711
40000  40000  4       214.1110 214.1110

Residual checks PASSED

End of tests

Sat 02/08/2014 
11:35 AM

Temps are still super-reasonable under this benchmark, but I'm also delidded. I'll see what 4.6 is later tonight and then post. It doesn't look like I should expect any voltage change for this test, though, so I think IBT is still good enough for determining stability. The optimized benchmark shows a lot higher GFLOPS though than IBT.

 

[BonzaiDuck]

Tristor's results are interesting (to me). 220 GFLOPS (given the problem-size) at 4.4? That's 90 better than my SB-K @ 4.7.

Its the same story over again. Back then some SB owners trashtalked IB. Now some IB owners trashtalk HW. All to feel better of having last gen tech I guess.

Some of us were probably put off by the extra trouble with de-lidding and the mixed results with it.

The die-shrink may have taken us into "new territory." Who am I to say? But I have a physicist-friend who is confined to his bed with MS. He had been predicting this matter of "leakage" for a couple years.

Also, the frequency distribution of "over-clockability" seems "all over the map." I'm wondering -- in addition to the TIM and IHS issue -- if Intel is finding it harder to QC the assembly-line output, or if this is again a result of the die-shrink.

CAn't argue with the performance, though .. .

 

[Tweakin]

...I've got friends with six-year-old dual core systems who believe the technology has far exceeded their needs or perception of progress.

I agree!!! If it not for the fact that I needed to replace my Q9650 as I gave it to my son, I can't imagine how long I would have been happy with that machine running 4 cores @ 3.2. It did everything I asked to to and then some!

Honestly though, giving the machine to him was an excuse for me to tinker again...

 

[rtsurfer]

Tristor's results are interesting (to me). 220 GFLOPS (given the problem-size) at 4.4? That's 90 better than my SB-K @ 4.7.

Interesting.

Although, Turn HT off, you will see better scores in Linpack that way.

I want to see the actual difference in performance between SB and HW.

 

[BonzaiDuck]

I agree!!! If it not for the fact that I needed to replace my Q9650 as I gave it to my son, I can't imagine how long I would have been happy with that machine running 4 cores @ 3.2. It did everything I asked to to and then some!

Honestly though, giving the machine to him was an excuse for me to tinker again...

Parallel experiences! I just wish I'd kept the modded case I built for the Q6600 before I gave it to my brother. It would've been a water-cooler aficionado's dream.

About this "excuse to tinker." I was going to start another thread: partly a joke, but it still has a serious dimension to it.

"There should be a forum entitled "12-step program for obsessive over-clockers and compulsive builders."

There's this news-item, released two months ago:

http://www.maximumpc.com/intel_release_8-core_haswell-e_desktop_processors_third_quarter_2014

I've almost finalized a decision to build an IB-E system this year, and wasn't planning to finish it until well into the summer. But even the details boil down to the infancy of the new chipset and memory spec: "DDR4?? I don' need no stinkeen DDR4's! Whadda I need stinkeen DDR4's for?! . . . "

We're all Lemmings, and Intel is like the Pied-Piper of Hamlin . . . or Haswell . . . whatever . . .

Although, Turn HT off, you will see better scores in Linpack that way.

I think you have that right. And like another thread I created about "troubleshooting hardware versus drivers," there was this aspect that the additional layer of complexity I added was using the system to run my "home theater" gear with Media Center (and . . . "HDCP"). When I'd disable HT, I was suddenly blocked from my "premium channels" and DVR playback because of . . . . Jack Valenti's legacy . . . So I left it enabled . . .

Even so -- I could turn it off, run the LinX tests and see what happens. Then turn it on again. Contrary to popular opinion, I don't run stress tests while Media Center is active. But I probably could, I guess . . .

 

[Gikaseixas]

4.6 with 1.29 voltage, i've run this thing @4.7 but was unstable in games and some apps

 

[BonzaiDuck]

4.6 with 1.29 voltage, i've run this thing @4.7 but was unstable in games and some apps

Anybody know where the "elbow" is in the exponential map of voltage versus Mhz on those Haswell suckers?

Although, Turn HT off, you will see better scores in Linpack that way.

I just remembered, on this angle. IDontCare had posted a guide about running stress-tests on HT-enabled processors. The idea was to set "affinity" of the stressing program to four of the eight threads. This would then give max. GFLOPS reading per iteration. While this made the readings consistent across iterations, do you have any further insight that the overall GFLOPS would improve? I think the point of IDC's explanation was that tweaking the stress program as described would still provide the true measure.

 

[Tweakin]

There's another thing we haven't covered in this thread. Overclocking results are significantly affected by cache speed and RAM speed. Unless stated otherwise, we should probably assume overclocks are at default cache clocks and DDR3-1600.

Good point! As you can see by my sig, I run 4.3 on the core and 4.0 on the cache. If I drop the cache down, I can achieve a better OC but I still have to dump way too much voltage into the chip.

 

[rtsurfer]

I just remembered, on this angle. IDontCare had posted a guide about running stress-tests on HT-enabled processors. The idea was to set "affinity" of the stressing program to four of the eight threads. This would then give max. GFLOPS reading per iteration. While this made the readings consistent across iterations, do you have any further insight that the overall GFLOPS would improve? I think the point of IDC's explanation was that tweaking the stress program as described would still provide the true measure.

I was also posting based on IDONTCARE's thread.

Setting linX to run on 4 threads gives us better results because when you set it at 8 it stress the CPU as if it has 8 real cores.

While we have 4 real + 4 Logical on our CPUs. In most real world applications HT gives us benifits but when you stress the CPU to max with something a.k.a Linpack, that can actually use that many cores at their full potential then you end up with a penalty.

This HT penalty was discussed heavily in AMD's Bulldozer threads as AMD uses something called CMT, while Intel uses SMT.

 

[Tweakin]

..."There should be a forum entitled "12-step program for obsessive over-clockers and compulsive builders."

LOL...I would be a charter member for sure.