Overclock.net › Forums › Intel › Intel CPUs › Haswell Overclocking Guide [With Statistics]
New Posts  All Forums:Forum Nav:

Haswell Overclocking Guide [With Statistics] - Page 590

post #5891 of 19578
Ya uncore really is just for squeezing that extra 0.1%-0.5% if you are that hardcore. As Darkwizzie says you should NEVER sacrifice Core mutl for Uncore (or even RAM though RAM has a slightly larger impact on performance). "Core is king!"

Here are some numbers I"ve gotten using two other different benchmarking apps, and much more aggressive changes in Uncore.

Core x46
RAM x1600 (stock)

Uncore (aka CacheRatio, aka Ring): x35
RealBench: 650-660 range (multiple runs)
XTU: 1038 (consistent number after a few runs)

Uncore (aka CacheRatio, aka Ring): x44
RealBench: 660-663 range (multiple runs)
XTU: 1041-1046 (multiple runs)

As you can see bumping Uncore by 900Mhz made a very small difference:
  • XTU -- taking the largest delta -- it accounts for a mere 0.77%
  • RealBench -- taking the avg delta since the ranges are relatively broad -- 0.99%
You will never be able to tell that difference.

NOTE: I like RealBench because it's not as synthetic as the others, it basically runs a script of GIMP, does some x264 and then runs a large video in the background while compressing and uncompressing files. It aims to simulate a more "real world" use pattern.
XTU basically is a modified version of Prime95 as far as I know.

DRAM overclocking made a larger difference

Core: x46
Uncore: x44
DRAM: 2133
RealBench: 677 (two runs, didn't test more)
XTU: 1118-1128

Improvement over x46/x44x/1600
RealBench: 3.35%
XTU: 8.1%
These results make sense since in a synthetic Prime95 bench (where you're not doing everything in-place, ie you're using RAM to shuttle data) memory speed is important as it is the bounding factor (after CPU speed), whereas in a more realistic real-world usage pattern such as RealBench you have a lot of disk access etc (I'm running on an SSD but it's still significantly more expensive to talk to that + the SATA interface + the long trip to memory/CPU etc).

So if anything, this shows (everything just confirms what Darkwizzie says)
1) Core is king and you should always optimize first and foremost
2) RAM
3) Uncore

Caveats & Disclaimers:
* one test I didn't do is if the RAM impact on scores would be the same if Uncure was at stock x35. I dont' think we would see a significant difference in results, but there could be cross-effects and convexity between the variables so something to bear in mind. If I'm not too lazy I can try it out a bit later if people are interested.
* I didn't set either XTU or RealBench priorities to "Realtime" in TaskManager which in retrospect was probably a mistake. That could potentially account for the small inconsistency in scores I got, but I was careful to not have anything else running, I could rerun the tests if people insist.
* This is specific to MY system. Your system may be different. However, I think this does generalize and reflects a tendency for Haswell.
Edited by creos7 - 11/25/13 at 9:23am
post #5892 of 19578
Quote:
Think it's low because of stability?

No, it's just hard to get points. I run a pretty lean system without windows aero etc, also tuned RAM quite a bit.

Download Maxxmem, make sure that your 2600 is significantly better scoring than the 2400 XMP, because raising clockspeed will make all of the timings that you don't set loosen automatically, and there's like thirty of them. The mistake is often made where people lose performance by clocking higher - a few Maxxmem runs should tell the tale if it's set to high priority and you use the same CPU settings while you're testing RAM (changing CPU clocks will make results better/worse too, so you only want to change RAM when you're testing)

After that.. It's just a few percent away. You got 971 right out? Maybe that's 980 in safe mode. Maybe with uncore up, it's 985. RAM tweaks? 990. Add 1mhz baseclock and you're there. Quite hard to do in reality, but stacking small things on top of eachother can go a surprisingly long way (:

1k is just a super epeen thing, i'd love it. Same with 200 single core. Not much practical use, as long as you're relatively close. I've beaten some peoples legit 4.8 with my 4.6 scores (partially because of RAM etc), and that makes me laugh a bit
Insert Name Here
(14 items)
 
  
CPUMotherboardGraphicsRAM
6700k Asus Maximus VIII Hero Gigabyte Aorus Xtreme 1080ti Corsair LPX 2x8GB 3200c16 
Hard DriveHard DriveCoolingOS
Old Seagate HDD Samsung 850 EVO Thermalright Silver Arrow SB-E SE Windows 7 Home Premium 64 bit 
MonitorKeyboardPowerCase
Asus PG258Q (240hz + Gsync) WASDKeyboards.com v1 semi custom w/ mx browns, ... Superflower Golden Green HX550 Air540 
MouseMouse Pad
Logitech G Pro Qck+ 
  hide details  
Reply
Insert Name Here
(14 items)
 
  
CPUMotherboardGraphicsRAM
6700k Asus Maximus VIII Hero Gigabyte Aorus Xtreme 1080ti Corsair LPX 2x8GB 3200c16 
Hard DriveHard DriveCoolingOS
Old Seagate HDD Samsung 850 EVO Thermalright Silver Arrow SB-E SE Windows 7 Home Premium 64 bit 
MonitorKeyboardPowerCase
Asus PG258Q (240hz + Gsync) WASDKeyboards.com v1 semi custom w/ mx browns, ... Superflower Golden Green HX550 Air540 
MouseMouse Pad
Logitech G Pro Qck+ 
  hide details  
Reply
post #5893 of 19578
Quote:
Originally Posted by Cyro999 View Post

No, it's just hard to get points. I run a pretty lean system without windows aero etc, also tuned RAM quite a bit.

Download Maxxmem, make sure that your 2600 is significantly better scoring than the 2400 XMP, because raising clockspeed will make all of the timings that you don't set loosen automatically, and there's like thirty of them. The mistake is often made where people lose performance by clocking higher - a few Maxxmem runs should tell the tale if it's set to high priority and you use the same CPU settings while you're testing RAM (changing CPU clocks will make results better/worse too, so you only want to change RAM when you're testing)

After that.. It's just a few percent away. You got 971 right out? Maybe that's 980 in safe mode. Maybe with uncore up, it's 985. RAM tweaks? 990. Add 1mhz baseclock and you're there. Quite hard to do in reality, but stacking small things on top of eachother can go a surprisingly long way (:

1k is just a super epeen thing, i'd love it. Same with 200 single core. Not much practical use, as long as you're relatively close. I've beaten some peoples legit 4.8 with my 4.6 scores (partially because of RAM etc), and that makes me laugh a bit

Cyro, you seem to know a fair amount about RAM etc. I have MemTweakIt which doesn't work to tweak stuff from the OS (says because of win8.1 limitations, not sure how true that is), but it does display the current settings. THere is a "DRAM Efficiency Score" there. I get a score of 33760 when I use the XMP 2133 profile on my GSkill Trident X. Is that a reasonable guide of memory performance or you woud recommend Maxxmem?

I have to say I've been having a fair amount of trouble getting my system stable with 2133... I've had to upp SA DIO AIO a fair amount (to about 1.15 SA AIO and 1.25 DIO) and enable 110% DRAM Current Capacity and Extreme DRAM Phase and DRAM Voltage to 1.62 (even though it's rated at 1.6V). I was stable at 30 Passes on x264 trying now OCCT but before the latest changes it gave an error after 3 hours (I usually consider 8hours as a stable run). I'm rerunning now with the latest changes, we'll see how it goes.

Is it better to relax the CAS for higher speed, or on the contrary? I'm on an ASUS Hero so any pointers are super welcome. DRAM has been a total time sink frown.gif
post #5894 of 19578
^I don't actually know that much about RAM.

Stability can be troublesome sometimes. I'd say, set 4ghz core or stock settings, validate RAM stability, and then go back to your 24/7 overclock and see if your system is still stable with RAM up - if not, make adjustments afterwards. Lower RAM speeds til it works then tighten timings again, or even drop core clock a little if you see issues with RAM all the way down to like 1600mhz (it can happen)

A few of your questions are answered by Maxxmem. It's just a quick click 10-15 second RAM benchmark that will throw out some numbers, higher = better. They're consistent if you keep the same CPU settings and show performance improvements or regresses, big or small, you can compare across frequencies with them - etc 2133 cas9 vs 2400 cas10. GL
Edited by Cyro999 - 11/25/13 at 9:58am
Insert Name Here
(14 items)
 
  
CPUMotherboardGraphicsRAM
6700k Asus Maximus VIII Hero Gigabyte Aorus Xtreme 1080ti Corsair LPX 2x8GB 3200c16 
Hard DriveHard DriveCoolingOS
Old Seagate HDD Samsung 850 EVO Thermalright Silver Arrow SB-E SE Windows 7 Home Premium 64 bit 
MonitorKeyboardPowerCase
Asus PG258Q (240hz + Gsync) WASDKeyboards.com v1 semi custom w/ mx browns, ... Superflower Golden Green HX550 Air540 
MouseMouse Pad
Logitech G Pro Qck+ 
  hide details  
Reply
Insert Name Here
(14 items)
 
  
CPUMotherboardGraphicsRAM
6700k Asus Maximus VIII Hero Gigabyte Aorus Xtreme 1080ti Corsair LPX 2x8GB 3200c16 
Hard DriveHard DriveCoolingOS
Old Seagate HDD Samsung 850 EVO Thermalright Silver Arrow SB-E SE Windows 7 Home Premium 64 bit 
MonitorKeyboardPowerCase
Asus PG258Q (240hz + Gsync) WASDKeyboards.com v1 semi custom w/ mx browns, ... Superflower Golden Green HX550 Air540 
MouseMouse Pad
Logitech G Pro Qck+ 
  hide details  
Reply
post #5895 of 19578
Quote:
Originally Posted by Cyro999 View Post

^I don't actually know that much about RAM.

Stability can be troublesome sometimes. I'd say, set 4ghz core or stock settings, validate RAM stability, and then go back to your 24/7 overclock and see if your system is still stable with RAM up - if not, make adjustments afterwards.

A few of your questions are answered by Maxxmem. It's just a quick click 10-15 second RAM benchmark that will throw out some numbers, higher = better. They're consistent if you keep the same CPU settings and show performance improvements or regresses, big or small, you can compare across frequencies with them - etc 2133 cas9 vs 2400 cas10. GL

thanks.
post #5896 of 19578
Can a stable OC get unstable when i use RAM which has a better CAS latency than the one before...(same RAM frequency =2133mhz)
post #5897 of 19578
If the RAM is stable and working fine at stock CPU clocks, i'd expect there to be little or no change in IMC stress for the same frequency RAM
Insert Name Here
(14 items)
 
  
CPUMotherboardGraphicsRAM
6700k Asus Maximus VIII Hero Gigabyte Aorus Xtreme 1080ti Corsair LPX 2x8GB 3200c16 
Hard DriveHard DriveCoolingOS
Old Seagate HDD Samsung 850 EVO Thermalright Silver Arrow SB-E SE Windows 7 Home Premium 64 bit 
MonitorKeyboardPowerCase
Asus PG258Q (240hz + Gsync) WASDKeyboards.com v1 semi custom w/ mx browns, ... Superflower Golden Green HX550 Air540 
MouseMouse Pad
Logitech G Pro Qck+ 
  hide details  
Reply
Insert Name Here
(14 items)
 
  
CPUMotherboardGraphicsRAM
6700k Asus Maximus VIII Hero Gigabyte Aorus Xtreme 1080ti Corsair LPX 2x8GB 3200c16 
Hard DriveHard DriveCoolingOS
Old Seagate HDD Samsung 850 EVO Thermalright Silver Arrow SB-E SE Windows 7 Home Premium 64 bit 
MonitorKeyboardPowerCase
Asus PG258Q (240hz + Gsync) WASDKeyboards.com v1 semi custom w/ mx browns, ... Superflower Golden Green HX550 Air540 
MouseMouse Pad
Logitech G Pro Qck+ 
  hide details  
Reply
post #5898 of 19578
I'm still in progress of overclocking, but the best I could do it seems is 4.4GHz core clock @1.39V (XTU 3 Hours....) ....which seems like a very high voltage for such a low clock. Not sure if it is worth it over 4.3GHz @ 1.26V (barely verified).

I should add heat was not an issue, with the temps bouncing around between 60-70C withj 4.4GHz @ 1.39V for 3 hours, stability seems to be my issue. If the temperatures are under 80C with XTU for 8+ hours, and the voltage is under 1.5V I should be fine?? Or will the voltage alone lead to dramatic life reduction of the CPU? I realize though that once I am done OCing, I will set the voltages to be adaptive so in reality the overvolting will be intermittent... but still I am unsure how hard I want to push the voltage even if the temps are good... given I want the CPU to last 4-5 years at least.
Edited by KillerDJK - 11/25/13 at 11:30am
post #5899 of 19578
Reading through what I miss, the result with core, uncore is pretty much what I got as well. Only thing for me was I reach a point when going higher did not give better number and in fact was lower in certain instances. I been playing with my chip and see how it run with my everyday use. I am happy to report like Dark, I been running it at 4.6 ghz with no issue, not a single bsod after 3 full week. Over the weekend I been playing with 4.7 and so far no issue. I got it to run at 4.8 but the jump in voltage is pretty high and haven't found stability yet, it will crash in games and rendering. All of this is without stress testing but real world usage. These are just quick test to see if I can get higher multi without having to change any other value, beside vcore of course, I know I am stable at.

So my chip is scaling like this:
4.4 need 1.25 volt. But I wanted 4.5 so majority of my testing was at 4.5. I needed 1.26 volt for 8 hours stable and 1.265 to be 12 hours stress xtu and 10x x264 stable.

4.6 at 1.36 volt and 4.7 at 1.365. This is without test. I am sure if I had to start stressing them I would probably fall into line with the 4.4/4.5 result and need a bit more vcore to pass stress test.

I was able to boot up at 1.39 for 4.8 but so far upping vcore up to 1.45 still can't find anything stable, either crash or freeze. But I notice that I am getting 101 error instead of 124 once I passed 1.42 volt.

Something I found interesting was, I left dram volt to auto to see what it was pulling and hitting 4.8 it say instead of 1.664 I needed 1.68 for my dram voltage. I manually set it to 1.665 with no issue so far.

So does this mean higher overclock need more dram voltage as well as cache voltage given my 101 error at 4.8? Is there a point when more voltage in general is needed across the board due to high overclock?

It seem that with more time and fine tuning and of course better cooler I might get more out of this chip. My real limiting factor at this moment is heat.

I will add that these quick setting without spending the time to do extensive testing and optimizing my benchmark number are not too difference. I see improvement in benchmark number yes but not the number I am reading when going up a multiplier.

Bench real quick with xtu,cinebench and geekbench, beside 4.8 ghz no bench because it crash like a crash test dummy wheee.gif
ThermalTron
(16 items)
 
  
CPUMotherboardGraphicsRAM
4790k MSI Z87 Mpower MSI 7870 Team Xtreem Dark 
Hard DriveHard DriveCoolingOS
Western digital Samsung 840 pro H100 Windows 8 64bit enterprise 
MonitorMonitorMonitorKeyboard
Apple HCD Dell LG Corsair K90 
PowerCaseMouseMouse Pad
Kingwin 850 Platinum Thermaltake Chaser MK1 TT Theron Roccat Sense Glacier Blue 
  hide details  
Reply
ThermalTron
(16 items)
 
  
CPUMotherboardGraphicsRAM
4790k MSI Z87 Mpower MSI 7870 Team Xtreem Dark 
Hard DriveHard DriveCoolingOS
Western digital Samsung 840 pro H100 Windows 8 64bit enterprise 
MonitorMonitorMonitorKeyboard
Apple HCD Dell LG Corsair K90 
PowerCaseMouseMouse Pad
Kingwin 850 Platinum Thermaltake Chaser MK1 TT Theron Roccat Sense Glacier Blue 
  hide details  
Reply
post #5900 of 19578
Quote:
Originally Posted by KillerDJK View Post

I'm still in progress of overclocking, but the best I could do it seems is 4.4GHz core clock @1.39V (XTU 3 Hours....) ....which seems like a very high voltage for such a low clock. Not sure if it is worth it over 4.3GHz @ 1.26V (barely verified).

I should add heat was not an issue, with the temps bouncing around between 60-70C withj 4.4GHz @ 1.39V for 3 hours, stability seems to be my issue. If the temperatures are under 80C with XTU for 8+ hours, and the voltage is under 1.5V I should be fine?? Or will the voltage alone lead to dramatic life reduction of the CPU? I realize though that once I am done OCing, I will set the voltages to be adaptive so in reality the overvolting will be intermittent... but still I am unsure how hard I want to push the voltage even if the temps are good... given I want the CPU to last 4-5 years at least.
ouch frown.gif i'm sorry, that is a bit on the poor-peformer side ya. I need 1.39 to hit 4.7, i'm broken-in solid 4.6 @1.28.
This is with your RAM and Uncore/CacheRatio/Ring at stock (i.e. manually set to x34)? Have you tried lowering RAM to 1333? (or 1300 whatever that standard is) to see if it allows for lower vcore?
New Posts  All Forums:Forum Nav:
  Return Home
  Back to Forum: Intel CPUs
Overclock.net › Forums › Intel › Intel CPUs › Haswell Overclocking Guide [With Statistics]