I have tested a bit more. I previously had my computer stable at 150BCLK, 29x CPU (4350MHz), 10xSPD(1500MHz) 6-6-5-15-1T, Uncore 20x (3000MHz), QPI 36x(5400MHz), DRAM 1.600V, IOH Core 1.100V, QPI/Vtt 1.275V, LLC enabled and a DVID +0.11250V, resulting in max load (and max idle) voltages of 1.344-1.360V.
(Edit: new paragraph) I previously did a couple of tests
with LLC enabled and LLC disabled and posted my findings, concluding that LLC enabled gave lower idle voltages (when you adjust the voltage accordingly) by 0.032V, but higher IBT VH load voltages by 0.016V. However - I had only checked stability with IBT VH, and hadn't checked stability with neither Prime95 nor IBT at Maximum, and hadn't enabled EIST nor C-states either, and this last part caused the idle voltages and CPU multiplier to run at full tilt all the time. How would LLC enabled/disabled work out with EIST/C-states enabled, and how would Prime95 and IBT Max stability be? I was sure I would have to go up with DVID to get IBT Max and Prime95 stable, but would load and idle voltages (for the lowest stable DVID voltage) be different depending on the LLC settings? Sure - the maximum idle voltage was higher with vDroop (that's the nature of the beast), but how would the idle voltages be when the CPU clocked down into lower power states and correspondingly lowered the voltage? That's what I wanted to find out.
So, I have tested a bit more. I have previously spent some time trying (and failing) to get the bloody IMC and RAM stable with all six sticks at 1600 (and the rated 1866) MHz, but discovered that I could get it seemingly stable at 1500 MHz with only 1.500V DRAM and a QPI/Vtt of 1.195V. Woot - lower voltages means lower temperatures.
However, after running several tests, both with and without LLC, I found out that the voltage settings weren't Prime95 stable - I found out that my voltages were juuust a smidgen too low on QPI/Vtt. But after a bump up to a QPI/Vtt of 1.215V I was back in business - the LLC enabled settings were finally stable. I was now able to run new tests, making a true apples to apples comparison - try to find out which voltages and temps I would get with LLC disabled (and thus vDroop enabled) vs having LLC enabled (and vDroop disabled).
I first ran this setup (with LLC enabled) with a DVID of +0.07500V (1.312-1.328V), but IBT (Intel Burn Test) failed at VH (Very High), so I ramped up to a DVID of +0.09375 (giving voltages of 1.328-1.344V), but Prime95 failed after 7hrs 34 minutes. So - I ramped up again to DVID +0.11250, and it was finally stable - Prime95 ran successfully for 11hrs 30minutes. (I knew this DVID voltage was going be stable; I had previously run the previous stable setup with QPI/Vtt 1.275V and DRAM 1.600V for 31 hours - it was now the CPU voltage that was the limiting factor.)
So - the setup
(Edit: Hardware): Gigabyte GA-EX58-UD5, Xeon W3690, 6x2GB Corsair Dominator GT (7-8-7-20-2T 1866, if you care).
I used the following settings:Frequencies: 150BCLK, 29x CPU (4350MHz), 10xSPD(1500MHz) 6-6-5-15-1T, Uncore 20x (3000MHz), QPI 36x(5400MHz)
Voltages: DRAM 1.500V, IOH 1.100V, QPI/Vtt 1.215V, LLC enabled, DVID +0.11250V
C1, C3/C6/C7, EIST and Hyper Threading are all enabled (I'll reboot to BIOS and verify that later).
When switching LLC off/on, the only
changes are the DVID voltage adjustment and the LLC Enabled/Disabled.LLC ENABLED:
So - I started the computer. Idle voltages were 1.088-1.360V
I ran IBT
(Intel Burn Test v2.54) at Very High
x10, giving load voltages of 1.344-1.360V
and temperatures of 68/68/63/63/69/68 °C
(and GFlops of 81.5-81.7).
As you can see, both the idle max and load max are at 1.360V - both idle and load voltages are 1.344-1.360V, due to the LLC being enabled. And - because I have EIST and the C-states enabled, the CPU often clocks down and goes down to a floor of 1.088V.LLC Disabled:
I then disabled LLC. When I disable LLC, load voltage is going to be a lot lower with the same idle voltage, and thus I knew I would have to raise the voltage adjustment.
I started turning DVID up one notch - at +0.13125. This gave an idle voltage of 1.104-1.376V. I fired up Prime95, figuring that it would probably be the limiting factor, and sure enough, I got a BSOD 0x50. Up with the DVID.
I tried a DVID of +0.15000V, with idle voltages of 1.120V-1.392V. I ran Prime95 for 10 minutes, with Prime95 voltages of 1.328-1.344-1.360V. It ran OK, so I fired up 10x IBT VH, for load voltages of 1.328-1.344V - and it passed! I went on to Prime95, and it ran 19 hours 22 minutes stable (yaay!). But - when I fired up IBT at Maximum (10068MB) it failed already on the second go. So - the voltage would have go up even one more notch.
I then went for a DVID of +0.16875V
, giving idle voltages of 1.136V-1.408V
. I ran IBT Maximum (9936MB)
x10, giving load voltages of 1.360-1.376V
and temps of 66/66/63/62/67/67 °C
As you can see here, the load voltages are markedly lower than the max idle voltages - that's how vDroop works.
Edit II: pictures of the completed run - GFlops scores of 57.9-58.4 GFlops:
To check the max temperatures and load, I ran IBT VH x10
, giving load voltages of 1.344-1.360V
and temps of 68/68/64/64/69/70 °C
(and GFlops of 81.3-81.7) - essentially the same voltages and temps (and GFlops) I got with my stable settings of LLC enabled.
Again we see that when IBT has finished, the max voltage jumps back up from 1.344-1.360V and up to 1.408V.
But - would my previous configuration with LLC enabled also be stable in IBT maximum?LLC Enabled pt II:
I reenabled LLC
and reduced the DVID
back to +0.11250V
. Idle voltages were 1.088-1.360V, with several dips down to 1.072V. I fired up IBT Maximum (10068MB), and as always with my LLC (at a given DVID) I got load voltages of 1.344-1.360V
. The temps were 65/66/63/63/67/67 °C
- the same as I got with LLC disabled.
..Oh yes it was stable.
Edit II: With GFlops scores of 57.4-58.2.Conclusion
So - there you have it, apples to apples: On my computer and my CPU, overclocked to 4350MHz, I need the same load voltages at the most extreme load (with IBT VH) to get my three tests (IBT VH x10, IBT Max x10 and Prime95 for many hours) stable, both with LLC enabled and LLC disabled - I get the same kind of GFlops score, the same temperatures and the same IBT VH load voltages.
But - this also means that the idle
voltages differ. The idle voltages were 1.136-1.408V with LLC disabled (and vDroop enabled), but only 1.088-1.360 (occasionally 1.072) with LLC enabled - a whopping 0.048V (to 0.064V) difference! And - the IBT Maximum (and presumably Prime95) ran mostly at one notch higher with LLC disabled - at 1.360-1.376V, versus 1.344-1.360V with LLC enabled.
Edit II: The IBT Max GFlops scores were slightly
higher with LLC disabled than with LLC enabled - this could be to circumstances (I have several times experiences higher/lower scores of up to +/- 1 GFlops on different reboots with the same settings), or it could be because the load voltage is slightly higher with LLC disabled during IBT Max runs.
So - you get the same load voltages when running the very most extreme test (IBT VH), but when idle or at lower loads, disabling LLC will give higher voltages - up to a difference of 0.048-0.064V at idle. And that is quite a bit!
So - certain forum members claims that enabling LLC will cause voltage spikes of up to 0.05V. Does it? And if so - is it dangerous? I don't know. I want to find out, though.
I do know, however, that for my overclock on my CPU and my motherboard I can choose between enabling LLC, and maybe risk potential spikes of 0.05V when going down to a lower idle load - or disable LLC, and get a voltage 0.05V higher at idle all the time
Do I want a potential extra spike of 0.05V, or do I want a surefire constant idle voltage delta of +0.05? For me that question is simple to answer. Your mileage may vary - your answer may be different than mine, or it may be the same. Or - maybe your CPU and/or motherboard behaves differently. I do not know - I have only got this one X58 motherboard. Warning: Spoiler! (Click to show)
When Intel has stated that the max voltage for my CPU is 1.4V - do I want to run my CPU at 1.344-1.360V, which maybe may get an occasional spike up to 1.4V - or do I want my idle voltage to frequently and for longer periods go up to 1.408V - all the time? I'll leave it up to the reader to guess as to what my own answer to this particular dilemma might be.
But anyway - when enabling LLC - do remember to lower your voltage at the same time
- the same idle voltage will give a higher load voltage when LLC is enabled, and thus you must reduce it to get the same load voltage.And thus - as long as you adjust the voltage accordingly when enabling/disabling LLC (for my overclock it was a difference of 0.048V measured in Windows and 0.05625V chosen in the BIOS), enabling LLC will give you lower voltages.
(Edit: specifying obvious conclusion:) And yes, due to the danger of potential voltage spikes, I will of course
tolerate higher idle voltages with LLC disabled than with LLC enabled. Idle voltages of 1.408V with LLC enabled? I don't think so!
Not on my rig, anyway! Higher idle voltages than 1.408V with LLC disabled? Nope. Don't want it. Even 1.408V is a bit too high.
So - any comments/questions to (the results of) my little test?
Some background information:
- When Windows has started, my motherboard and software reports CPU vCore voltages in steps of 0.016V. As the BIOS has voltage steps of 0.00625V, I usually go up and down three of those BIOS steps at a time, giving steps of 0.01875V, which is the closest I can get to the 0.016V measured steps. If I find that a step of 0.01875V takes me two 0.016 steps up when checking the voltages in Windows, I usually reboot and try a smaller step instead. But - this happened not to be the case in the tests above.
- On my computer, IBT at VH (Very High) generates the most amount of heat - it generates more heat than IBT at Maximum. So - that's why IBT VH gives the lowest voltages with LLC disabled and generates the highest temps.
- I have found that IBT gives lower GFlops scores when I have just ran Prime95 for many, many hours, so in my tests I always boot after a Prime95 run, to keep all the scores and results more consistent.
- Both CPU-Z and EasyTune6 report the correct voltages, but they have different polling intervals. When taking screenshots I have tried to get shots of one showing the lowest idle voltage and the other one showing the highest idle voltage, to get both in several screenshots. The voltages do vary all the time when idle...