Prime stabilty - is it always vcore?

eddiew · 8 Jul 2018 at 21:38

Hey peeps,

Not looking to reopen the old "it's stable except Prime crashes" (because Prime has proven that a number can get calculated wrong and therefore something else will eventually throw an error up too).

But the thing is, I've been gaming all day at ~1.29-1.31v without a wobble, and for all I know it will go lower yet and still be "game stable". But non-AVX Prime fails if vcore is below 1.334. Which is still fine for 5ghz I guess, and it's running in offset mode so it gets nice and low at idle. No complaints, really

But my question to the experts; is there anything besides vcore that influences Prime and might be a better place to tweak? I'm slightly suspicious of the memory, which is running XMP at 3200 C14, and is 16gb per stick and therefore dual ranked and relatively hard work for the memory controller. Currently at 1.35V.

I'm probably quibbling about a voltage tweak that's only worth a couple of degrees C, if that, but I'd be interested to know if there's anything I can use to keep vcore down, if only for academic reasons

Bugbait · 9 Jul 2018 at 10:26

With higher memory clocks (usually 3600+ but can be 3200+ with tighter timings) you often need to add some voltage to VCCIO (memory controller) and VCCSA (system agent). As per Vcore, just add a bit at a time.

eddiew · 9 Jul 2018 at 12:20

Bugbait said:
With higher memory clocks (usually 3600+ but can be 3200+ with tighter timings) you often need to add some voltage to VCCIO (memory controller) and VCCSA (system agent). As per Vcore, just add a bit at a time.

I'd read elsewhere that VCCIO at 1.05v is usually fine - do you have a suggestion for a baseline to start at?

eddiew · 9 Jul 2018 at 18:09

Well I read around a bit and set both to 1.15 which seems completely safe, although they bump up a tad under load. VCore is -100mV offset, which reads 1.328 in HW monitor under load, and never passes 1.344.

30 minutes into Prime 95 and it hasn't passed 80 degrees, in a system that is being very quiet in a warm room. Initial results are optimistic

Update: had a BSOD around 32 minutes

Have loosened my memory timings from 14 to 15, if that stabilises it then I guess I may need to look more at VCCIO than VCore. It's possible that 16gb sticks are just really hard to handle. Currently 51 minutes into a blend (but we all know they can fail after 12 hours when they want to

)

Meantime this is my HWMonitor if anyone has any advice:

I dunno what the TMPIN3/5/6 are. The hottest is within 1 degree of the CPU all the time so I suspect it's reading the same sensor twice. Maybe #5 is a VRM?

zipp0r · 10 Jul 2018 at 00:13

its better to test cpu overclock separately from playing with memory speed/timings. eg set ram to a real low speed to completely rule it out.
small fft is for cpu testing. large ftt is for ram/memory controller.
so pass 24hrs of small before moving to the large. dont even bother with blend.

eddiew · 10 Jul 2018 at 08:04

zipp0r said:
its better to test cpu overclock separately from playing with memory speed/timings. eg set ram to a real low speed to completely rule it out.
small fft is for cpu testing. large ftt is for ram/memory controller.
so pass 24hrs of small before moving to the large. dont even bother with blend.

Yeah, I feel like I've been fighting on two fronts here...

That said, 3 hours 20 of Prime last night before my PC went to sleep (forgot it would do that). 1.2 vccio and vccsa (reads 1.232 in hwmonitor). I guess those dual ranked c14 dimms are punishing the memory controller pretty hard. Maybe I could drop one of those two settings a notch, but that's an experiment for another day and I doubt it will change anything tangible.

VCore is -80mv offset, showing as 1.312 under gaming loads, and 1.344 with spikes to 1.360 while stress testing. I'm bemused as to why the two are different when both are 5ghz... maybe my LLC is a bit aggressive? But if so that's great, because I know this thing will game for days at the lower voltages without a problem! If it's taking itself up by that extra smidge only when under high load, that sounds like the ideal outcome

eddiew · 10 Jul 2018 at 20:51

Spoke too soon... Sigh. VCCSA up to 1.264 and still bombed out on me this aft.

Right, well, memory down to 2666 to remove a variable. Read up on VCCIO and pretty sure it doesn't really have an impact on this overclock; busted that back to 1.1v, which helped thermals what with the creeping VCCSA. Wouldn't be surprised if it can go lower still, but I'll just try to get a working solid baseline first.

1 hour 25 of small FFTs so far... temps the same as yesterday.

Oi oi oi, I must admit I had hoped for a slightly smoother experience from a binned chip