The Reality of Haswell Overclocking - Results Poll

[BonzaiDuck]

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

Long is the road, hard is the path, and so many among the afflicted . . .

Thinking of Phillip Seymour Hoffmann and the flurry of CNN spotlights on the burgeoning Heroin Addiction problem . . . . Heroin Addicts don't have Newegg Preferred Accounts with "deferred payment options" . . . .

 

[Tristor]

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

When you say 4.6, are you saying it's stable in your applications or stable in a stress test? I really wonder, because that's an incredibly low voltage for 4.6. Chip lottery and all that, but my VID of 1v chip delidded w/ AIO water can't pass IBT at 1.29, but boots just fine at that voltage at 4.6GHz.

 

[Termie]

Really interesting results so far. Looks like with over 40 data points, the average is around 4.35GHz.

Some posts have argued that these aren't reliable, but you could also make the argument that some people haven't actually pushed their chips all the way. I for one have never tried more than +0.05v as I'd probably need water cooling to use much more than that.

Anyway, my take on this is that Haswell isn't really all that much worse than IVB, maybe 100-150MHz lower headroom, and about 300-350MHz lower than SB. That still makes Haswell the fastest of the chips.

 

[tviceman]

With my delidded 4770k, I'm at 4.2ghz @ 1.187 24/7, hitting around 70-71 C in IBT (very high) with a middle-of-the-road tower cooler.

 

[mike5757]

I have my 4670K running at 4.1 GHz with 1.21 V on air (Hyper 212 EVO). I tried to go for 4.3 GHz, but I just couldn't make it happen. Either I have a poor overclocker, or I am missing something.

When I tried for 4.3 GHz, I left the uncore at 38x and RAM at stock speed and voltage. I kept upping the voltage all the way to 1.31 V, but it still wasn't stable. At 1.31 V the CPU clock started fluctuating, sort of like it was being throttled, but temps were in the low 80s. I found it really odd that other people reported much higher temperatures with voltages as high as I was going. Although, my computer room was seriously cold, 15.5 C or 60 F, but I doubt that accounts for the whole difference.

Still a heck of an improvement over my old E6600. I wish I would've had it before I converted 11 or so seasons of TV shows from DVD with HandBrake.

 

[BonzaiDuck]

I have my 4670K running at 4.1 GHz with 1.21 V on air (Hyper 212 EVO). I tried to go for 4.3 GHz, but I just couldn't make it happen. Either I have a poor overclocker, or I am missing something.

When I tried for 4.3 GHz, I left the uncore at 38x and RAM at stock speed and voltage. I kept upping the voltage all the way to 1.31 V, but it still wasn't stable. At 1.31 V the CPU clock started fluctuating, sort of like it was being throttled, but temps were in the low 80s. I found it really odd that other people reported much higher temperatures with voltages as high as I was going. Although, my computer room was seriously cold, 15.5 C or 60 F, but I doubt that accounts for the whole difference.

Still a heck of an improvement over my old E6600. I wish I would've had it before I converted 11 or so seasons of TV shows from DVD with HandBrake.

You might as well just add the difference from your own in "prevailing room ambient" from a decent sample of members or users -- to estimate what the load value average would be in a warmer situation.

 

[DigDog]

^ lucky. i followed the same guide, have a CM Nepton 280 rad, and can't get it stable past 4.3Ghz.

 

[Tristor]

I'm doing testing at 4.6 with the optimized linpack and the results are... discouraging... to say the least. I've had to raise my vCore to 1.385 so far, am about to run the test again. I get about 90% of the way through and then I BSOD with "clock watchdog timer". I'm wondering if it may be my memory holding me back because I've made a point of doing all my testing and overclocking with my memory set to XMP-1 (DDR3-1866 CL9 @ 1.5v). I'm considering raising vCSA to see if it makes any difference. I feel like I'm right on the cusp of stability but not quite getting it. My temps are also not quite to my liking, peak 84C @ 1.380vCore.

Also of note, all my testing has been done with my cache speed at 40x (4000MHz) with a 1.300vCache set.

I'll hopefully update this post with an edit shortly with good Linpack results at 4.6. I was going to see if I could get it to 4.8 under 1.45vCore, but with the AVX voltage jump I'm not sure it's safe to shoot for.

I can certainly say the optimized Linpack is more stressful on the stability of the CPU than IBT though. Remember, I have been IBT stable as my 24/7 OC since the other day at 1.370vCore at 4.6GHz.

 

[24601]

I'm doing testing at 4.6 with the optimized linpack and the results are... discouraging... to say the least. I've had to raise my vCore to 1.385 so far, am about to run the test again. I get about 90% of the way through and then I BSOD with "clock watchdog timer". I'm wondering if it may be my memory holding me back because I've made a point of doing all my testing and overclocking with my memory set to XMP-1 (DDR3-1866 CL9 @ 1.5v). I'm considering raising vCSA to see if it makes any difference. I feel like I'm right on the cusp of stability but not quite getting it. My temps are also not quite to my liking, peak 84C @ 1.380vCore.

Also of note, all my testing has been done with my cache speed at 40x (4000MHz) with a 1.300vCache set.

I'll hopefully update this post with an edit shortly with good Linpack results at 4.6. I was going to see if I could get it to 4.8 under 1.45vCore, but with the AVX voltage jump I'm not sure it's safe to shoot for.

I can certainly say the optimized Linpack is more stressful on the stability of the CPU than IBT though. Remember, I have been IBT stable as my 24/7 OC since the other day at 1.370vCore at 4.6GHz.

Testing with cache at slower speed than core defeats the purpose of Intel Math Library Optimized Linpack.

That's why I try to get people to "show their work" with the actual GFLOPS results

 

[Termie]

I'm doing testing at 4.6 with the optimized linpack and the results are... discouraging... to say the least. I've had to raise my vCore to 1.385 so far, am about to run the test again. I get about 90% of the way through and then I BSOD with "clock watchdog timer". I'm wondering if it may be my memory holding me back because I've made a point of doing all my testing and overclocking with my memory set to XMP-1 (DDR3-1866 CL9 @ 1.5v). I'm considering raising vCSA to see if it makes any difference. I feel like I'm right on the cusp of stability but not quite getting it. My temps are also not quite to my liking, peak 84C @ 1.380vCore.

Also of note, all my testing has been done with my cache speed at 40x (4000MHz) with a 1.300vCache set.

I'll hopefully update this post with an edit shortly with good Linpack results at 4.6. I was going to see if I could get it to 4.8 under 1.45vCore, but with the AVX voltage jump I'm not sure it's safe to shoot for.

I can certainly say the optimized Linpack is more stressful on the stability of the CPU than IBT though. Remember, I have been IBT stable as my 24/7 OC since the other day at 1.370vCore at 4.6GHz.

I would caution you against using those voltages for anything other than benchmarking. That's not a 24/7 overclock.

 

[Tristor]

I would caution you against using those voltages for anything other than benchmarking. That's not a 24/7 overclock.

I'm aware. After this I actually intend to lower my overclock to 4.4 for 24/7 use because it's at a significantly lower voltage (that's where I hit my "wall" where things went somewhat exponential) and because the performance difference is minimal.

Testing with cache at slower speed than core defeats the purpose of Intel Math Library Optimized Linpack.

That's why I try to get people to "show their work" with the actual GFLOPS results

It's not really something you have a choice about though, it's simply not going to be stable at a 4600MHz cache speed. Not happening. I wasn't able to get 4400MHz at 1.330vCache, and I refuse to go any higher on vCache because I consider those voltages to be of higher risk than vCore. Considering the default speed of the cache is 3400MHz and I'm running at 4000MHz, it's still 600MHz faster than it otherwise would be and less of a delta with the core.



All that said, it seems it's irrelevant. Using the optimized linpack I am unable to get stable at 4.6 at voltages I would even remotely deem acceptable for that clock speed. I can keep ramping voltage, but I noticed an irregularity that concerns me so I'm going to postpone further benchmarking until I can analyze that further. Here's the data I have so far:

4.6@1.385vCore
Peak 76/84/83/78 (this is the irregularity btw, the core-to-core delta should be much smaller considering I've delidded and have lapped my IHS and waterblock, I think I may have somehow not gotten enough CLU on in a spot or the IHS is very not flat on the underneath inside the area that sits over the die).

Intel(R) Optimized LINPACK Benchmark data

Current date/time: Sun Feb 09 06:19:32 2014

CPU frequency:    4.598 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      0.017      40.0328  1.083189e-012 3.693953e-002   pass
1000   1000   4      0.011      63.0760  1.083189e-012 3.693953e-002   pass
1000   1000   4      0.005      122.7537 1.083189e-012 3.693953e-002   pass
1000   1000   4      0.005      123.6466 1.083189e-012 3.693953e-002   pass
2000   2000   4      0.040      134.0015 4.220901e-012 3.671667e-002   pass
2000   2000   4      0.040      134.8228 4.220901e-012 3.671667e-002   pass
2000   2000   4      0.039      138.4833 4.220901e-012 3.671667e-002   pass
2000   2000   4      0.038      138.8564 4.220901e-012 3.671667e-002   pass
3000   3000   4      0.123      146.0860 1.016483e-011 3.914231e-002   pass
3000   3000   4      0.114      157.3810 1.016483e-011 3.914231e-002   pass
3000   3000   4      0.115      157.2500 1.016483e-011 3.914231e-002   pass
3000   3000   4      0.114      158.6057 1.016483e-011 3.914231e-002   pass
4000   4000   4      0.253      168.7408 1.906425e-011 4.155234e-002   pass
4000   4000   4      0.252      169.4226 1.906425e-011 4.155234e-002   pass
4000   4000   4      0.251      170.2759 1.906425e-011 4.155234e-002   pass
4000   4000   4      0.257      166.3512 1.906425e-011 4.155234e-002   pass
5000   5000   4      0.477      174.7770 2.299338e-011 3.206242e-002   pass
5000   5000   4      0.470      177.4455 2.299338e-011 3.206242e-002   pass
5000   5000   4      0.470      177.3550 2.299338e-011 3.206242e-002   pass
5000   5000   4      0.471      176.9655 2.299338e-011 3.206242e-002   pass
10000  10000  4      3.264      204.2947 9.420734e-011 3.321846e-002   pass
10000  10000  4      3.257      204.7249 9.420734e-011 3.321846e-002   pass
15000  15000  4      10.476     214.8195 2.137378e-010 3.366406e-002   pass
15000  15000  4      10.481     214.7155 2.137378e-010 3.366406e-002   pass
20000  20000  4      23.507     226.9138 3.723286e-010 3.295924e-002   pass
20000  20000  4      23.502     226.9643 3.723286e-010 3.295924e-002   pass
25000  25000  4      46.011     226.4214 5.678848e-010 3.229357e-002   pass
25000  25000  4      46.048     226.2407 5.678848e-010 3.229357e-002   pass

During these runs the highest GFLOPs I saw was 231, however in all tests around the time it hits the 30000 or 35000 problem size I end up getting the watchdog timer BSOD.

 

[24601]

I'm aware. After this I actually intend to lower my overclock to 4.4 for 24/7 use because it's at a significantly lower voltage (that's where I hit my "wall" where things went somewhat exponential) and because the performance difference is minimal.



It's not really something you have a choice about though, it's simply not going to be stable at a 4600MHz cache speed. Not happening. I wasn't able to get 4400MHz at 1.330vCache, and I refuse to go any higher on vCache because I consider those voltages to be of higher risk than vCore. Considering the default speed of the cache is 3400MHz and I'm running at 4000MHz, it's still 600MHz faster than it otherwise would be and less of a delta with the core.



All that said, it seems it's irrelevant. Using the optimized linpack I am unable to get stable at 4.6 at voltages I would even remotely deem acceptable for that clock speed. I can keep ramping voltage, but I noticed an irregularity that concerns me so I'm going to postpone further benchmarking until I can analyze that further. Here's the data I have so far:

4.6@1.385vCore
Peak 76/84/83/78 (this is the irregularity btw, the core-to-core delta should be much smaller considering I've delidded and have lapped my IHS and waterblock, I think I may have somehow not gotten enough CLU on in a spot or the IHS is very not flat on the underneath inside the area that sits over the die).

Intel(R) Optimized LINPACK Benchmark data

Current date/time: Sun Feb 09 06:19:32 2014

CPU frequency:    4.598 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      0.017      40.0328  1.083189e-012 3.693953e-002   pass
1000   1000   4      0.011      63.0760  1.083189e-012 3.693953e-002   pass
1000   1000   4      0.005      122.7537 1.083189e-012 3.693953e-002   pass
1000   1000   4      0.005      123.6466 1.083189e-012 3.693953e-002   pass
2000   2000   4      0.040      134.0015 4.220901e-012 3.671667e-002   pass
2000   2000   4      0.040      134.8228 4.220901e-012 3.671667e-002   pass
2000   2000   4      0.039      138.4833 4.220901e-012 3.671667e-002   pass
2000   2000   4      0.038      138.8564 4.220901e-012 3.671667e-002   pass
3000   3000   4      0.123      146.0860 1.016483e-011 3.914231e-002   pass
3000   3000   4      0.114      157.3810 1.016483e-011 3.914231e-002   pass
3000   3000   4      0.115      157.2500 1.016483e-011 3.914231e-002   pass
3000   3000   4      0.114      158.6057 1.016483e-011 3.914231e-002   pass
4000   4000   4      0.253      168.7408 1.906425e-011 4.155234e-002   pass
4000   4000   4      0.252      169.4226 1.906425e-011 4.155234e-002   pass
4000   4000   4      0.251      170.2759 1.906425e-011 4.155234e-002   pass
4000   4000   4      0.257      166.3512 1.906425e-011 4.155234e-002   pass
5000   5000   4      0.477      174.7770 2.299338e-011 3.206242e-002   pass
5000   5000   4      0.470      177.4455 2.299338e-011 3.206242e-002   pass
5000   5000   4      0.470      177.3550 2.299338e-011 3.206242e-002   pass
5000   5000   4      0.471      176.9655 2.299338e-011 3.206242e-002   pass
10000  10000  4      3.264      204.2947 9.420734e-011 3.321846e-002   pass
10000  10000  4      3.257      204.7249 9.420734e-011 3.321846e-002   pass
15000  15000  4      10.476     214.8195 2.137378e-010 3.366406e-002   pass
15000  15000  4      10.481     214.7155 2.137378e-010 3.366406e-002   pass
20000  20000  4      23.507     226.9138 3.723286e-010 3.295924e-002   pass
20000  20000  4      23.502     226.9643 3.723286e-010 3.295924e-002   pass
25000  25000  4      46.011     226.4214 5.678848e-010 3.229357e-002   pass
25000  25000  4      46.048     226.2407 5.678848e-010 3.229357e-002   pass

During these runs the highest GFLOPs I saw was 231, however in all tests around the time it hits the 30000 or 35000 problem size I end up getting the watchdog timer BSOD.

Thank you for that data .

Now I have a baseline to be able to better compare other people's overclocks with.

 

[tviceman]

With my delidded 4770k, I'm at 4.2ghz @ 1.187 24/7, hitting around 70-71 C in IBT (very high) with a middle-of-the-road tower cooler.

Also want to add that I had to go from 1.085v @ 4.0ghz to my current setup (4.2ghz @ 1.187v) to get stable in both IBT and games (bioshock infinite would BSOD me even if I passed a 5 loop very high IBT run with any less voltage). I could not get 4.4ghz stable without pushing voltage, temps, and power beyond my acceptable limits.

My older 2600k ran at 4.5ghz 24/7 and never got above mid 60's in anything.

 

[Termie]

I've finally got some screenshots, which the OP originally requested:

Note the temp and voltage difference when running IBT versus running Crysis 3. This is without changing any settings - I ran Crysis right after IBT, only closing HWMonitor to clear the min/max records. This is using a Cooler Master downdraft cooler (S524).

 

[rtsurfer]

Testing with cache at slower speed than core defeats the purpose of Intel Math Library Optimized Linpack.

That's why I try to get people to "show their work" with the actual GFLOPS results

Stop making your own rules and contradicting eveything else out there.

You can check multiple Haswell overclocking guides on Youtube frim Linus, JJ, etc. And even the guide in OCN where people have said that Cache speed doesnt make a huge performance impact.

You are better off trying to hit a higher Vcore than to try to get a 1-1 Vcore to Cache ratio.

You will need a real cherry Cpu to hit 1-1 Vcore and Cache at high multipliers.

 

[Tweakin]

...no screen shot as I really can't be bothered...

thanks

 

[WhoBeDaPlaya]

Couldn't get much past 4.55GHz with all three chips I have (well, not with a reasonable amount of V anyway).

 

[Tristor]

I've finally got some screenshots, which the OP originally requested:

Note the temp and voltage difference when running IBT versus running Crysis 3. This is without changing any settings - I ran Crysis right after IBT, only closing HWMonitor to clear the min/max records. This is using a Cooler Master downdraft cooler (S524).

Thanks for posting screenshots. You may want to know that to get a reliable stress test out of IBT you need to set the cores equal to your real cores (i.e. disregard HT) and set RAM usage to maximum as per the sticky at the top. At 4.4, you're getting 10 less GFLOPs than I am in in IBT on an i5-4670k@4.4.

 

[24601]

Stop making your own rules and contradicting eveything else out there.

You can check multiple Haswell overclocking guides on Youtube frim Linus, JJ, etc. And even the guide in OCN where people have said that Cache speed doesnt make a huge performance impact.

You are better off trying to hit a higher Vcore than to try to get a 1-1 Vcore to Cache ratio.

You will need a real cherry Cpu to hit 1-1 Vcore and Cache at high multipliers.

Keep subzero cooling.

I'm sure that'll be very useful data for people in general.

I'm sure "stable" obviously means "didn't crash yet, making sure not to do anything strenuous, as strenuous isn't reality."

Keep lying to yourself and others, that's what makes these lists pointless in the first place. Just like OCN runs.

 

[Kenmitch]

I sold my 4670k the other day.

My chip was Linpack11 stable at 4ghz 1.06v fixed vcore. I was running this as my official 24/7 overclock.

It would do up to 4.8ghz with some tweaking in bios. 4ghz was the sweet spot as going to 4.2ghz was the 1st larger vcore bump requirement.

I think it crapped out around 4.8ghz or so. Guess it may have been the MB's power phases that limited the chip....Will never know for sure.

This is the highest clocked I currently have a picture of. Was using a H100i so didn't worry too much about temps for the most part. No delid but would have required it for 24/7 use unless one can handle the H100i in ear bleed fan mode.

 

[Termie]

Thanks for posting screenshots. You may want to know that to get a reliable stress test out of IBT you need to set the cores equal to your real cores (i.e. disregard HT) and set RAM usage to maximum as per the sticky at the top. At 4.4, you're getting 10 less GFLOPs than I am in in IBT on an i5-4670k@4.4.

Well, I gave that a shot, but with just four threads, my temps and my GLOPs were lower. I believe those instructions are out of date. In my experience, you actually have to shut off hyperthreading to get it to stick to four threads.

Anyway, seems stable to me, and I don't use IBT as a benchmark, just a stress test. With HT on, the temps are higher, so that's how I run it.

 

[dragantoe]

I ran my 3570k at 4.7 ghz 24/7, does haswell provide a significant clock for clock performance increase? or is it dumb for someone to choose haswell over ivy bridge for a new system?

 

[Termie]

I ran my 3570k at 4.7 ghz 24/7, does haswell provide a significant clock for clock performance increase? or is it dumb for someone to choose haswell over ivy bridge for a new system?

The only thing that doesn't make sense is to upgrade from IVB to Haswell.

But yes, Haswell does perform better overall (on average 9%), and of course makes sense for a new build assuming you get an overclock that is on par lottery-wise. That means about 100-150MHz lower than IVB.

A 3570K@4.7 is way above average, and I wouldn't buy Haswell assuming you'll get a 4.5-4.6 overclock.

 

[Tweakin]

...It's not really something you have a choice about though, it's simply not going to be stable at a 4600MHz cache speed. Not happening. I wasn't able to get 4400MHz at 1.330vCache, and I refuse to go any higher on vCache because I consider those voltages to be of higher risk than vCore. Considering the default speed of the cache is 3400MHz and I'm running at 4000MHz, it's still 600MHz faster than it otherwise would be and less of a delta with the core.

That's as high as I could run my cache at also...I'm back down to stock voltage on the cache at 4.0Ghz. This might be a common wall...

 

[tviceman]

The only thing that doesn't make sense is to upgrade from IVB to Haswell.

But yes, Haswell does perform better overall (on average 9%), and of course makes sense for a new build assuming you get an overclock that is on par lottery-wise. That means about 100-150MHz lower than IVB.

A 3570K@4.7 is way above average, and I wouldn't buy Haswell assuming you'll get a 4.5-4.6 overclock.

Honestly I think upgrading from decent Sandy Bridge OC'd chip (like I did) doesn't make much sense either unless the PCIe 2.0 bus is keeping you up at night.