cekim wrote:
1. you are running at 2133 memory, so I am not sure this applies, but something about the latest bios hates 1T memory timings
-- even if they worked before... they don't now, but curiously, 2T runs nearly as fast as 1T used to when you benchmark it
2. 240% should not be required, 140 should be more than enough for anything non-LN2
3. Power Phase Response? I usually end up with that back on Standard, but results may vary.
4. You will see FAR, FAR more improvement with faster memory than you will going from 4.5 to 4.6.
-- just checked your system and those are 2133 CAS15 dimms. Your system will show you a nice bump in performance (again far, far bigger than 4.5 to 4.6 will) if you upgrade to some faster memory.
-- it depends on what you do, but if you run anything that is large enough to spill out of the cache significantly, you will see 1:1 speedup there with memory clocks
5. You will see FAR, FAR more improvement with faster cache than 3GHz vs shooting for 4.6.
-- 4GHz should be easy with .100 to .250 offset voltage (depending on your setup)
6. I always seem to end up with LLC5 to get things rock-solid
7. 32G of memory usually isn't a big strain, but I generally end up with a +0.200 offset to SA as well.
8. Get aida64 stable first, but encoding will stress it more in my experience (don't bother with the others). Aida seems a bit weak on the memory sub-system. I had things rock-solid in aida that would fail linux stressapp immediately and reliably with anything less than 20-30mV more than aida's stable values.

Lynkdev wrote:
Ok ill switch back to 2T for memory. All the system is for is gaming, no encoding or anything.
Ill switch back to 140%. I set it to power phase response just following someone else's power settings.

How do i set faster cache if i decided to stay at 4.5? better memory?
Does cpu input voltage matter at all for adjusting?

cekim wrote:
"Max Cache Ratio"

Set that to 40,41,42,43,44,45 as your system tolerates. You will likely find it requires exponentially more votage beyond 41/42, but this various system by system.

I'd start with a 40x cache ratio with an offset of +0.200 and see how that goes. You are in "full manual" mode, so I don't think you'll have offset as an option. In that case start at 1.0v and see how it does. The cache represents a LARGE portion of the chip in terms of area and power consumption, so you will have to gauge both stability and heat production as your limiting factors. Less of an issue until you get above 1.2v where it starts to show itself rapidly.

Lynkdev wrote:
Thanks,

I went back to experiment with XMP. where is this offset?

cekim wrote:
In order to see/use offsets, you have to disable "Fully Manual Mode". Which I generally do. I usually start by applying XMP, reverting BCLK back to 100, manually entering all my memory settings (because XMP gets them wrong) and then start playing with offsets and adaptive voltages. You can't use adaptive for the cache, but you can and should use it for the CPU.

eric.iannone wrote:
If my power plan is set to high performance will adaptive voltage still work?

Sent from my SM-N920V using Tapatalk

cekim wrote:
Yes, no, and maybe?

Yes, it will still function
No, windows will peg your frequency at 100% and voltage along with it
maybe:
a. if you are running linux rather than windows, even the "perf" governor will throttle the clock under low load.
b. even if you disable "fully manual" mode, you can still set a fixed voltage if you like, but its really a waste IMO.

I'll say this, broadwell shows near zero cost to me running in "balanced" mode. Anything I've benchmarked, showed no difference whether synthetic or real-world application.

Haswell-E does show a measurable impact running in balanced mode.

HOWEVER, the difference is small and you are dumping a lot of heat and power to do nothing. So, even Haswell, I generally have in a balanced mode.

Arne Saknussemm wrote:
All you have to do is edit the power plan...

In advanced settings set "minimum processor state" to 10% and speedstep works perfectly...