Originally Posted by owcraftsman
Originally Posted by afkingjay
with your knowledge should i change my LLC, Phase Control, and Duty Control and CPU Current Capability or keep it at default settings? I see alot of guys doing this. Just not sure what this does, and when i Google this its not in lamens terms so I'm not sure what changing this would do... this is my First O.C+ LLC Medium (25%)
+ 0.040 offset[/B]
thanks a bunch
System specs as follows
corsair vengeance ram at xmp 1866
hyper 212 evo cooler
right now im at 100x4.7 epu disabled and offset =.035
LLC = stock
Phase Control = stock
Duty Control = Stock
CPU Current Capability = 100%
Most OC's above 4.6 to 4.8 folks will set:
LLC to Ultra High,
Phase Control & Duty Control to Extreme,
& CPU Current Cap to 130%
the offset looks fine however it never hurts to see if it's stable at a lower offset if not possible you can always go back.
I have some of my setting listed on the spread sheet linked in my sig feel free to try what you see there and add your own.
Keep in mind I have the 2600k but they OC essentially the same only real difference is Hyperthreading and sometimes require a bit more vCore
If you are stable with the settings you listed that would be fine too however I suspect not.
FYI: My definition of stable is running Prime95 custom utilizing max memory for 10 hrs min which is how long it takes to run all available test within Prime.
I have 8GB memory installed and set Prime to use 6154GB memory sometimes I have to select a bit less to get it to run it all depends on how much the system is using at the time you start the test.
Certain test will raise the heat high just like Linx but other test stress lots of memory etc etc making Prime95 the best overall indicator of a stable system if you can pass all the test.
Your advice is excellent, however I disagree with your recommendation of LLC and I explain why in detail below in each of my iterations of testing.
We also have different definitions of stability. Ten hours is not
going to run every iteration of FFT within Prime95 if you use the 15 minute value for each test. The more tests you run, the more FFTs it will test. I began to compile a list using a Blend customized for 90% RAM use (which on my system is closer to 6975 MB memory being used) and I ran that Blend for 1 minute cycles. I got bored after 90 minutes and had recorded 89 distinct FFTs and had not yet reached any duplicate FFTs. If you extrapolate that out to the default 15 minutes per test, you'd have to run for 22.25 hours and you'd still not have a duplicate FFT. You would run through each type of FFT and probably get a statistically relevant sample but you're not going to hit them all until you're at or beyond the 24 hour mark.
Also the Prime95 tests change depending on if you increase the maximum size of the FFT. (I did not when I tested this.)Each person's criteria for a stable system are different. Here are my criteria. (Click to show)
With all of that being said, my criteria for a "stable" system are 30+ iterations of Intel Burn Test on maximum (Linpack based, will give more thermal stress than Prime95 and will also force more Vdroop than Prime95 and give you a BSOD 0x0101 much faster than Prime95 will if your voltage is too low.) Following that, I immediately reset my temperature values in Real Temp (after recording them for reference) and I test a Prime95 blend customized to use the same amount of RAM Linpack was using. (This gives you a 95% use situation). I leave all other settings as the default Blend and I run this for at the very least 12 hours, but I prefer 18 or even 24 hours. This will catch BSOD 0x0124 memory over/under volt and/or cpu over/under volt that IBT is known to miss. After both tests, I will run a few PCMarks and 3DMarks and then just to make sure that I don't have any Northbridge or Southbridge controller issues that the other tests missed. If I've done any GPU OC I will run a FurMark (or OCCT or evga OCScanner, all use a FurMark variant) and stress the GPU as well as get a performance value from it to make sure that I'm not throttling at load.
And then I do a 24 or 48 hour folding run where I combine -smp #cores and a GPU fold simultaneously. It's actually possible to appear to be rock-solid stable on Prime95 and Linpack and yet still have errors due to OC. Folding@Home will catch these errors, especially GPU errors or chipset/controller errors that other benchmarks and stress tests aren't as likely to catch.
The final test is gaming and every-day use!
I also chose to deviate and use a 140% power profile rather than 130% and can justify it with numbers. (Note that the peak power consumed didn't change, but the computational efficiency improved.) Doing this requires that you reduce CPU PLL voltage values though. (Reduce PLL volt from 1.8 down to as low as 1.5, then raise power limit to 140% and very carefully increase PLL until your peak power hits just at or slightly below the power limit throttle mark. There's more detail on that later.
I have some general statements to make and then I will back them up with data from my latest (and rather exhaustive) experimentation with offset mode over-clocking.
My sig-rig has my components but I'll list them quickly here:
P8P67 WS Revolution
G.Skills Ripjaws (blue CL8) at XMP 1600 8-8-8-24-2 Two x 4 GB = 8 GB
Thermalright Venemous-X RT
I have similar specifications to the above poster. However, it is important to note, when looking at the temperatures I post, that my case and fan scheme are designed for the most quiet gaming possible in a rather outdated case that couldn't be any smaller without a micro ATX board design. My temperatures could be drastically improved in a more modern case with multiple 120mm and 140mm fans.
Here are my statements:
- The least amount of LLC to stay stable is the most thermally and computationally efficient setting you will get.
- The greatest amount of Current Capacity limitation you feel comfortable with is best to use. (I use 140%)
- The lowest peak Vcore and VID values that you can get without having BSOD 0x0101 is the best.
- The lowest Vdroop (minimum voltage at high load) that you can get BSOD 0x0101 or BSOD 0x0124 is the best.
- Enabling Internal PLL Overvoltage (no settings associated with it) will help with booting into windows for your first OCs in case there is a problem with your board.
- CPU PULL Voltage can be reduced to as little as 1.5000V and in fact, that's a good place to begin your OC at.
- Disabling LLC, C3 and C6 reporting, and using a manual voltage will allow you even more efficient over-clocks than offset, but you lose the ability to throttle your cpu down during idle or low-use scenarios.
You will have a more efficient (in terms of GFlops, Prime95 cycle times, TPF if you fold, or benchmark points on a bench) if you follow the above advice and in some cases you may find that a 100 MHz lower CPU clock will out-perform the next highest multiplier in terms of GFlops if you follow this procedure carefully. (In my case, the performance gain was significant.)
Here is my experience with pseudo-stable and final-stable OC's beginning with the stock system.It's important to note that absolutely none of these changes were tried at random in large increments, nor did I change multiple parameters at once. I'm simply condensing all of my trial and error into the combinations that resulted in the initial linpack/IBT testing stability while watching maximum Vcore and Tcore so as not to exceed design parameters. Each iteration listed here is anywhere from one to fifteen iterations in real experimentation to reflect the multiple change combination shown.Stock Settings. Always start here at stock, and change each feature you are considering adjusting here independently while observing its effect on temperature, voltage, and power consumption before attempting to change them at elevated clock/voltage levels. (Click to show)
I started with stock settings and I found that my motherboard and CPU together performed as follows:
0.960 Vmin (This was a baseline to determine absolute minimum stable voltage at 1600 MHz)
3700 MHz (Clock speed)
7090 MB RAM (amount of memory assigned for IBT during maximum run)
1.248/1.240 Vpeak (peak voltage in IBT / Vdroop at full load in IBT)
60/68/68/66 (Tcore max temperatures by core, lower is better obviously)
174.001 s (calculation time at that memory amount, is only relevant if memory used is identical)
102.68 GFlops (measure of efficiency of the system, higher is better)
All changes are made in bold face and are successive. meaning the changes in the previous test(s) apply to all subsequent tests as well, in other words, only new changes are noted. Values are not additive though, so a +0.25V does not reflect an addition of 0.25V onto the last value, but instead reflects that the value is +0.25V compared to stock.
Disabling LLC (Click to show)
+ 140% Temperature Capacity
+ LLC Regular (0%)
7130 MB RAM
As you can see, my thermal efficiency and voltages did not change, but my GFlops were reduced by a measurable amount. This change was to establish a baseline.
Changing to Current Capacity regulation (Click to show)
+ 140% CPU Current Capacity
7252 MB RAM
Note that by switching from Temperature to Current regulation for the Power Limits, my GFlops returned to stock values. Voltages have increased, but thermals have not, probably due to Heat-Sink/Fan efficiency.
Enabling Internal PLL Over Voltage to guarantee booting into windows. (Click to show)
+ Enable PLL O.V.
7235 MB RAM
No significant change here. This was to establish the effects of the changes on voltage in upcoming tests.
Reducing CPU PLL from 1.8 to 1.7 and adding a tiny fixed offset to prevent automatic offsetting. (Click to show)
- CPU PLL 1.7000 (reduced to value of 1.7 from stock/auto value of 1.8000V)
+ 0.005 Offset (total value 0.005)
7248 MB RAM
There are still no significant changes here, as I was attempting to establish a safe platform to begin my OC with. Voltages have increased some as have temperatures, but GFlops remain equivalent to stock values (as they should with a stock core clock.)
A stable 4.4 GHz OC using near-stock voltage and maintaining stock VID (VID undocumented.) (Click to show)
+CPU 44 (a departure from the stock 37)
7234 MB RAM
18.9% increase in CPU speed results in 15.9% GFlops. Note that Vpeak/Vdroop increased by 1.9% and 1.3% respectively. Thermals increased by 8.5% (maximum) over stock. This is an incredibly efficient overclock and many OC folks would stop right here, given that the 74C peak temp in IBT (the hottest test out there) is still 24C below TJMax and offers a huge margine of safety. Vcore is also a very comfortable value and the performance gain is quite close to the clock increase. The ratio of GFlops Gain vs Core Clock to acheive it is 84.1%
Jumping to 4.6 GHz OC and finding a slightly increased offset through trial and error. (Click to show)
+ 0.015 offset (total value 0.015)
7224 MB RAM
The results here indicate a 19.3% increase in GFlops at a 24.3% increase in clock speed. The ratio of GFlops gain vs core clock to achieve it is 79.4% Thermals increased over stock by 13.2% Voltages are 3.2% over stock.
Changing many parameters in a long string of iterations to reach 4.7 GHz IBT stable (Click to show)
+ CPU 47
+ CPU PLL 1.5000
+ 0.055 offset
7226 MB RAM
Note that I did not jump right to these PLL and offset changes. Moving from 4.6 GHz to 4.7 GHz was not nearly as easy as jumping straight to 4.4 GHz from stock or letting the automatic regulations hit 4.6 GHz stable with minimal effort. This was the result of about 20 iterations of successive testing in order to get IBT stable at the lowest possible voltage values. It's safe to say, given the sharp increase in voltage required to hit a 47 stable multiplier that I hit a small voltage wall here.
I have gained 21.7% in GFlops over stock at a 27% clock increase. The ratio of Gflops to clock to achieve it is 69.5% Voltages are 6.4% over stock. Thermals are 19.1% over stock.
Concluding that a 4.8GHz was not a wall worth pushing through in terms of temperature or voltage or in terms of the -reduction- in CPU computational power. (Click to show)
+ LLC Medium (25%)
+ CPU PLL 1.4875
+ 0.060 offset
+Spread Spect. Off (Spread Spectrum began to introduce random instability here and had to be disabled.)
7233 MB RAM
I went on to test 4.8 GHz to see what sort of performance gains I could expect and to see how big my voltage wall was that I hit. (I know that this chip is 5.0 GHz capable but I hit an even bigger voltage wall at the 4.9 and again at the 5.0 GHz mark and the throttling due to excessive power consumption was so bad, that I considered 4.8 GHz to be a practical limit in terms of performance efficiency.
So I have a 9.6% voltage increase a massive 33.8% thermal increase with a totally unacceptable temperature here. I can get that temperature and voltage lower by reducing the CPU Current Capacity, but them I will throttle the GFlops sooner. In fact, you can see that I have a reduced GFlops in this test as compared to the prior test. There's definitely a wall here and it's not worth trying to overcome it.
So I returned to my last known value of 4.7GHz and attempted to increase its efficiency...
Optimizing 4.7 GHz for maximum computational performance. Note the significant performance gain over the previous 4.7 GHz clock, based on -disabling- elevated LLC values. (Click to show)
- CPU 47
- LLC Regular (0%)
+ CPU PLL 1.50625
+ 0.060 offset
7199 MB RAM
1.328 Vpeak / 1.296 Vdroop
74C 84C 85C 81C
145.8 W peak
As you can see, I've managed to gain an extra 2 GFlops over the previous 4.7 GHz run at the exact same voltages. I did this by increasing PLL slightly and increasing the offset slightly, which had almost no measurable effect on voltages but did impact power consumption and thermals. Voltages are increased by 6.4%, thermals by 25%, CPU clock speed is the same 27% increase but now the GFlops increase is 23.84% My GFlops to core clock ratio improved to 88.3%
In terms of performance to clock ratio, this is about the best it gets with an OC and voltages are very reasonable. Although the thermal increase is also 25%, proportionate to my computing power (GFlops) increase, if you take into account that I'm allowing 145.8 W peak vs 95W peak reference, that's a 53.5% increase in electrical power consumption with only a 25% increase in thermal value. This is what separates a good heat-sink/fan from an inferior one. This is also the point where the water cooling guys come in and tell me how water cooling is even better. (And water cooling is better; I just couldn't afford it.)
The problem is that I'm only BSOD 0x0101 stable here (min Vcore) and I am still going to experience Prime95 Large FFT failures and folding inaccuracy and I'll crash in BF3 as well all with BSOD 0x0124 values due to insufficient Vcore at Vdroop levels or due to wildly cycling transients exceeding VID due to large offset and quick unload. I tried increasing the 350MHz phase count to a higher value but I wasn't any more stable and instead just consumed more electrical power and generated more heat.
Developing the 4.7GHz optimized profile above for maximum stability in a 24/7 environment, and concluding that for -my- system a tiny LLC was needed, but could be optimized for efficiency to -almost- as good as without. (Click to show)
- PLL Overvoltage Disabled (No need for it, as we boot into Windows (and Linux) just fine.
+ LLC Medium (25%)
+ 0.040 offset
7178 MB RAM
1.344 Vpeak / 1.320 Vdroop
75C 86C 85C 82C
145.8 W peak
26.5% thermal increase, 9.4% voltage increase, 27% clock increase, 23.5% GFlops increase, and my GFlops percent increase to Core percent increase is 87%. Compared to the last test, I'm slightly less efficient in terms of performance vs core, moderately less efficient in terms of temperatures, and substantially less efficient in terms of voltage. (50% less efficient than the previous test) This drastic reduction in voltage is not a very valid comparison though, as I'm
You'll notice that the peak voltage increased substantially over the last iteration as did the droop voltage. The increase to the peak voltage was a necessary evil in order to guarantee that Vdroop was >= 1.3200V. Any lower than 1.3200V on Vdroop and I would BSOD 0x0124. Now I could go with a LLC of 50% or 75% and bring Vpeak down to 1.328V while maintaining Vdroop >= 1.3200V but the problem with doing that is that my power consumption increased too rapidly with my 140% multiplier and I hit the current limit on the CPU, causing it to throttle down into the 120 to 122 GFlops range in IBT averages again. I could also reduce the CPU Current Limit to 130% or 120% and monkey with voltages but again, I had issues reaching the 126.5 to 127.0 GFlops range and in fact struggled to break 122.0 again.
So the above profile is what I settled on. It has almost as good of a peak GFlops as was previously recorded but is Prime95 Blend Customized for 97% RAM stable for 23 hours. (No screenshot for the sandy stable club updates as I didn't have enough memory to open the extra two or three instances of CPU-Z required and then print-screen. I tried, and I ran out of memory and got weird blank CPU-Z values until I closed the extra Windows and then the Real Temp were displayed again. In hindsight, I should have gone for only 92% memory not 97% memory. (I have my page-file disabled, so 8 GB is the max.)
Regardless, it's BF3 stable, 24/7 folding -smp 4 stable (on a 64 bit linux distro), IBT 25 iterations stable, and benchmark (Cinebench and PCMark) stable. I'm 3DMark stable too, not that a GPU bench should cause crashes, but it's possible.
My temperatures in BF3 are typically in the mid 50s to very low 60s with peaks at 65. (Temperatures are influenced by my GPU acting like a space-heater at this point.) In Prime95 they peak at 79C on the most intensive FFTs and remain a fairly constant 69 to 72C in most other instances. In 24/7 folding on the Linux distro, my peaks are currently 63/70/70/68 and this accomplishes my goal of being <= 70C for all daily operations. and to maintain at least a 10C margin of error from TJMax in the most aggressive heat-generating software out there. (IBT) I'll never hit Prime95 in-place small FFT or Linpack/IBT stress levels in the real world for more than a few seconds and even that's a stretch.
On a fixed voltage scheme, I can reach 5.0 GHz stable (albeit with high temperatures) and not have Vdroop issues. However, I don't like the idea of my core remaining in a higher voltage state 24/7 nor do I like the added fan noise or sensitivity to ambient temperature fluctuations.Edited by shad0wfax - 12/9/11 at 9:22am