Quote Originally Posted by Gr1mR34p3r View Post
@post #429

Great post, all of this is true, but I have something to add.
If you want the lowest load voltage and still be stable in light load (LinX 1 thread), you are best with LLC Regular.
If you want the lowest light load/idle voltage, then you would have to go with LLC Extreme, but this results in very high load voltage. So LLC Ultra High/Extreme is not so useful in my opinion.
Even LLC High requires around 0.04 (tested on 4.6GHz) more load voltage than LLC regular. And in light load (only 1 thread used) Regular requires only 0.01V more than High.
So by my experience, LLC Regular is the best, everything higher has more disadvantages than advantages.

And now that we know that LLC Regular's weakness is load voltage, best way to test it is Prime95 Small FFTs.
On LLC High, LinX 1 thread proved to be most stressful. Same goes for Ultra High/Extreme.

I've observed this effect enough to come back to this thread looking for answers. And it looks like I'm not alone.


So here's my conclusion on the problem. Feel free to correct me anywhere if I'm wrong.

The problem with the LLC is that the load voltages are different depending on your actual load on the processor.

With LLC on the load voltage is highest when all cores are under load.
Say that you're stable with 1.400v under 4 cores.
But when you reduce the load to say 1 core, LLC will pull back the voltage to like 1.200v. When that happens - instant BSOD. Reason: 1.200v is simply not enough for your current clock.

What you need is to hold 1.400v regardless of load. The only time it should be allowed to go down is when SpeedStep kicks in and actually lowers your frequency.

Over the past week or so, I was quite baffled at why my chip needed 1.400v to hold 4.6GHz stable over a 36 hour (not perfectly paralleled) task when it appeared to be prime and LinX stable at 4.7 GHz.
So I'll be playing around with lower LLC settings and higher voltage offsets.