Fine tuning a CPU overclock

Hey all,

So I wanted to ask your opinions on an overclock I've had for quite a while and made small tweaks to over time.

Currently I have a 4690k with an EVO 212, Asus Z97-E motherboard and 16 GB DDR3-1866 FuryX RAM.

After much experimenting, I first settled with a multiplier of 44 on all cores, at 1.250 vcore.

Later I heard that cache clock should be close as possible to core clock, so I boosted my cache to the same 44 at the same voltage. I found that this wasn't always 100% stable so I brought cache down to 42.

During this time I still had DDR3-1600 speed RAM, however later I upgraded to my current 1866 and set the XMP profile. The XMP profile however gave me occasional stability issues and so I boosted my vcore to 1.3.

This is completely stable, but I'm not sure if I have the best balance now. 1.3 vcore is really as high as I'm comfortable going, but do you think there's anymore tweaking I can do to improve things? It's not necessarily more speed I'm after, just better "balance".

I am no expert in terms of overclocking, I don't really mess with the cache multiplier or voltage.
1.3 is really high for 4.4GHz. I have a multiplier of 44 for 4.4GHz at 1.198V.
I also have a Kingston Hyper-X 1866 16GB kit.

When you add an XMP profile it is best to reset the overclock and start over. XMP changes a lot of things that can introduce unwanted variables into an overclock. I have no doubt you can maintain 4.4GHz on 1.25v with XMP, but just start over from scratch with XMP enabled and work your way back up.

I'm not sure what you mean by "work my way back up". The only things I changed were core clock, cache clock, vcore, cache voltage, and set RAM to XMP. I can't really see how it matters which order I set them in. Is there anything else I can tweak that might improve my stability?

I mean you will need to find the stability again. If you lost stability by adding XMP, than your current settings are not going to be good enough for 24/7. Starting over and moving in increments back to your current OC will help to find the stability again.

A good overclock requires more than just one step, you take things 100MHz at a time and adjust settings accordingly to keep each step stable. When you find the point which stability is lost and can't be simply fixed with small adjustments, you either bite the bullet and crank up settings to compensate or you revert to the previous stable step and call it a day.

You are in no way going for a world record OC here, so it should be easy to find that balance again and continue to enjoy the performance boost you gain. By starting over you may find a better balance as well, so instead of 1.3v you could end up around 1.22v or less, which is a lot better.

Well I have a good idea where my limits are, my question is what is the best balance? Is it better to have cache clock closer to the core clock but at the cost of other factors? Is it better for me to lose the XMP and go back to 1600 on my RAM so I can use lower vcore? Those are the kind of things I'm not sure about when it comes to overclocking.

  • Get your core clock as high as you can/want/temperatures allow without touching the uncore/keeping it at stock speed.
  • Get your memory as high as you can (XMP or manual OC) but don't sacrifice core clock for memory speed unless your program is one of those few workloads that actually benefit from RAM speed/higher bandwidth/lower latency.
  • Get your uncore as high as you can/want but never ever EVER sacrifice core clock or RAM speed for uncore speed. It doesn't affect performance by pretty much at all. If someone told you 1:1 core:uncore ratio is best, they're talking out of their arse.

IF your system isn't stable with XMP memory profile (with 1866 I seriously doubt that) it could be your IMC not able to keep up
You can try:
A) increase the, uhm, VCCIO voltage (IMC eg. memory controller voltage)
B) put the RAM back on manual and input the settings yourself but putting the RAM speed one notch slower.

The advise of "doing it again" after enabling XMP memory profile is valid but only really applies with ridiculously high speed RAM, read like 2666 instead of 1866 which is peanuts in comparison.
With high core OC + stupid fast RAM => "hitting" the IMC limits is much easier. Solution, like I said above is either increase IMC voltage or back the RAM speed down.

Edit: Another thing. Also, after you've found that happy medium with core/uncore/RAM, you can experiment with lowering PLL voltage in the hopes of
A) directly lowering your temperatures by a smidgen
B) indirectly lowering your temperatures by allowing you to drop your vcore down a tiny bit which will also bring your temperatures down a little bit.
It's one of those last things to do while fine tuning.

Remember to do plenty of stress testing with loads of different workloads and keeping and keeping an eye out for WHEA errors. So don't focus on P95 or IBT/Linx. Do stuff. Run FPU heavy workloads. Play some games. Run some encodes (HWBot x265 4K is nice). Rendering and/or raytracing. Or how about RealBench. And so on and so forth. Running just one or two different stress tests succesfully doesn't guarantee a properly stable system.
Just running a one hour x265 4K encode might surprise you by greeting you with a WHEA error or even a flat out crash.

Edit2: Aaand since this is Haswell, forget Linpack all together. That's just synthetic as fuck with AVX2.

Thanks for the detailed reply. That gives me some idea on the order to tweak in and what should have priority. For testing I firmly believe in real world applications vs burn testing. Funny enough I found assassins Creed unity and syndicate to be the ultimate test for me. If there's instability in my CPU, those games always seem to be the ones to find it.

Update: I reset my overclock and so far it looks stable at 4.4 and 1866 with a 1.250 vcore. I achieved this by leaving the cache clock at default this time as it was suggested to give this lower priority. If this overclock holds stable for a while I'll try pushing the cache clock to a more modest x40. Or I can leave it alone and try pushing to 4.5 GHz instead. Or I can leave both alone and see if I can bring my vcore down further to 1.230 or 1.240.