Originally Posted by Jcoffin1981
Originally Posted by shad0wfax
Try fiddling with your LLC, use offset to set your idle voltages where you want them and then configure your load voltages using the additional turbo voltage setting. Less LLC is always better, so long as you can keep it stable.
Ok, so I am currently at 5.1GHz and have been stable for months with zero issues. My LLC is set to "ultra," my load flucuates bet 1.376 and 1.384V, and I'm idleling at just under 1V. My offset is at -.025V. Is there really a benefit to decreasing my LLC and decreasing my offset? I'm not sure that if I don't have any idle BSOD issues that this should be messed with. What are your thoughts? After the foldathon I think I am going to work at 5.2GHz stable. Perhaps I'll try this technique.
We're digressing quite a bit from the Prime95 Stable aspect of this thread, so I'm going to hide this response in a spoiler. :) (Click to show)
You have a great CPU that's working wonderfully well with your system. Given that you're operating at lower voltages and temperatures at 5.1 GHz than I am at 4.9 GHz, the below may not apply to you and altering your LLC may not give you the same degree of change (or any change) in performance as I noticed with my system at your current frequencies.
When I tested with and without LLC I found that at 4.6, 4.7, 4.8, 4.9, and recently 5.0 GHz that the higher my LLC was set, the more thermal output I had at an equal load voltage and frequency, and the less efficiently my processor operated when compared to lower values of LLC.
I found this by running Intel Burn Test (Linpack) which puts about the hottest and most power-hungry load on the core possible. I ran 5 tests on "maximum" settings and recorded the minimum, maximum, and 5-run average Gflop/s values from each test at a fixed frequency and voltage. I then set that memory value via "Custom" to the same value that I ran the first test with, and tested in IBT again with varying degrees of LLC and different offset/turbo boost voltages to maintain stability. I held frequency constant at 4.6 GHz and ran with my Medium, High, and Ultra High LLC settings that gave me similar idle voltages and equal voltages at load for that frequency. I repeated this method, with the exact same "Custom" RAM to test value in IBT for 4.7, 4.8, 4.9, and 5.0 GHz. (I was not comfortable with the temperatures at 5.1 GHz so I aborted the test early.)
I found that, so long as voltage and frequency were held constant, and the memory amount was set to the same values in IBT that as LLC decreased, the Gflop/s increased. I'm sure that I could test the same in Cinebench and would see similar results. (Albeit the scores would be numerically closer together as Cinebench values are smaller numbers.) I have stable settings for my system at 4.9 GHz with Medium LLC. I can only get stable at 5.0 GHz on High LLC unless I'm willing to go insane with additional turbo voltage at 5.0 GHz with Medium LLC and suffer a significant Vdroop. I didn't feel comfortable with that, so I hit a wall between 4.9 and 5.0 GHz that required High LLC to get past. (Or it required more fine tuning than I was willing to put into it.) When I compared the High LLC 5.0 GHz results with my Medium LLC 4.9 GHz results, I found that the Gflop/s in IBT for both frequencies were identical; In my case, lower voltage, lower heat, medium LLC at 4.9 GHz gave me the exact same performance at an ultimate load point as higher voltage, higher heat, High LLC and 5.0 GHz did. So for my system, 4.9 GHz is better than 5.0 GHz (Or I need to spend quite a bit more time perfecting the 5.0+ GHz overclocks on my system.)
Clearly, you have a more capable system than I do, in terms of overclocking headroom, as your system is overclocking at lower voltages and temperatures than mine is, at that frequency. My guess is that you're probably farther away from thermal and current limiting states at 5.1 GHz than I was at 5.0 GHz, so you may not see as drastic of a change as I did. You may not even see any change in performance with LLC changes at all until you are at a higher frequency.
* It's important to note that IBT places an artificially high power and thermal load on a system and that in real-world computing, you will most likely never reach these operating parameters. Perhaps if you are folding -bigadv, performing heavy video encoding, or high-end rendering, you may come close to these values, but it's unlikely, as IBT is a worst-case scenario stress. That being said, the gflop/s efficiency improvements that IBT shows at "ultimate load" may not be present at 100% load in less floating point intensive operations. The reason I placed so much importance on this test is that I can have the confidence of knowing that I have the most efficient overclock possible, even in a worst-case load scenario, and that I will not be limited by any internal throttles under any circumstances in my daily operations.
Originally Posted by Gilvin
Thanks bro, that really clears my mind.
I'm now using 4.7G @ 1.384v at-load and 1.6G @ 1.080 off-load, offset +0.05 and LLC set to 2 (ASROCK Extreme4 Gen3), so i guess it is about 80% LLC, all the other settings are set by its default or auto.
I used to think that fix the offset voltage up and just fiddle the LLC level to get the correct working voltage, now it seems wrong, I'll try again to join the stable club, thanks.
My guess is that you can reduce your offset to -0.010 or even -0.030, leave your LLC alone, and then increase your turbo voltage a bit more to get the same target 1.384 V at 4.7GHz as you currently have. Note that you'll want to change the offset first, and increase the additional turbo voltage slowly so that you don't spike it by accident. Of course, your system may not like being under 1.000V at 1.6 GHz. (Mine runs at 0.920 V at 1.6 GHz, but yours may not.)
Keep tweaking, and good luck getting stable!
Edited by shad0wfax - 2/20/12 at 10:07pm