About a week ago, I decided to take another crack at OCing my 7700K, as I'd read a bit more on the subject and felt I could do better. I'd been running a 4.5GHz x4 clock for a while, and I had some time to kill (and more importantly, some stuff to procrastinate) so I gave it a go. With the multiplier set to 46 (so a 4.6GHz clock on all four cores, but SpeedStep still enabled) I got the voltage as low as 1.19 in the BIOS, which resulted in 1.136VCore under load. I was pretty happy with this - the ambient temperature is rather chilly in my room, but under synthetic stress tests it peaked at about 63C on a single core, with average temps during benchmarks and gaming loads significantly lower than that. Leaving it over night, I got idle temps down into the teens at times!
So, there's plenty of headroom in temperature and voltage, so I should be able to do better right? Partially. It's starting to look more and more like I just plain got a bottom-tier CPU, as the longer I play with it the less convinced I become that I'm still doing something wrong. I've been able to achieve low-confidence stability at the following settings (ambient temps approximately 17C, so not particularly scientific):
|Multiplier||Voltage (BIOS)||LLC||VCore (Load)||Stress Temp|
I was also originally experimenting with a higher LLC than I probably would have used in retrospect, as I was originally attempting to get VCore closer to my desired voltage, but I now understand that this can potentially cause spikes well above the intended voltage when the load goes away before the VRM can react - the complete opposite of what I want. For now I've left LLC at Level 1 (ASUS' lowest value) - when I decide on the clocks I want to run the chip at long term I'll investigate different levels' effects on my ability to lower core voltages further. I'm not entirely sure why the voltage under AVX loads is above the adaptive voltage, as LLC shouldn't be to blame at level 1.
I used OCCT for stability testing, as I've found it's rather good at quickly determining stability - no real need to run it longer than about 15 minutes, as in every test I tried if it didn't fail within the first 5 minutes or so, it would make it multiple hours. Furthermore, every configuration set I've abandoned the project at and gone to bed has worked fine for daily usage, gaming, etc, so I'm not really interested in leaving Prime95 go for 24 hours or anything stupid like that. I did also run a looped x264 stability test for two hours at each setting once I'd settled upon it, but I didn't find it particularly good at early detection unless I was a long way from stability. RealBench just plain sucks for stability testing, as for some reason my GPU driver doesn't like it (perhaps newer Nvidia drivers?).
What's interesting to note is that using the naive offset voltage method, even at 4x4.6GHz, the stable voltage is down beyond the reach of Voffset - that is to say, my CPU will chug along just fine at 4.6GHz at even lower core voltages (1.19v or so under load, maybe lower), but -0.170v is the point at which the voltage is too low at idle: the machine downclocks to 800MHz, the VRMs supply <0.5v, and the CPU hangs in short order. I was able to use "adaptive" mode to combat this, 0.56v at 800MHz seems to not crash ever at idle, but when using adaptive mode, the Vcore is never as low - you can see it sitting at 1.232, yet I've had this chip stable as low as 1.185v.
So, 5GHz? Not without a delid at least. I don't entirely understand how a delid helps stability (I would think it would help temperatures only), but allegedly it does, so there's still a chance. Playing around with the settings for several hours, and I can get Windows to boot at 5GHz with a VCore of ~1.4, but OCCT fails very quickly. 1.4V is substantially higher than I'd like to run at this point in my CPU's life (maybe when it was closer to being replaced) even were it stable. Since I'm running into temperature issues at email@example.com, I don't see 5.0 happening without a delid.
On AVX offsets: I originally played with a 200 or 300MHz AVX offset (which downclocks the CPU by that amount when executing AVX instructions, because they run quite hot and draw a lot of power), but I haven't found that necessary for stability purposes lately. 79C is some 60C above ambient, so my office would have to be incredibly hot before thermals become an issue at 4.6GHz. I'm optimistic I can lower the temps further with a delid, but worst-case scenario the AVX offset is there, slashing a few hundred MHz off the clocks under AVX loads is good for about 20C.
I think for now, 4.6GHz (maybe 4.7) is the sweet spot for me - I certainly don't need CPU power so much as to trade 200MHz for 12 degrees and 80mV. 4.7 might seem a goer, as the temperature and Vcore delta isn't terribly high - 4.8 is probably fine, but doesn't leave me much thermal headroom when summer arrives. "Stock" vcore on this board at 4.5GHz "turbo" (so one core) is around the 1.34v mark (!!!), so with any luck I'll lengthen the life of the CPU by almost as much of a factor as I'm sure I'll shorten it later in it's life.