System Voltages and stability

[Raimo5]

Hi all,

A new member here looking for a bit of advice from more veteran clockers. That being said, I am a long time consumer of ROG products.

I recently built a new system coming from i3-530, so it was a long time coming.

New PC specs:

CPU: i5-8600k @ 5.4GHz (will be replaced with 9700k when I get my hands on it)
Cooling: Corsair H150i PRO
MOBO: Maximus XI Hero
RAM: Corsair Vengeance LED 3200MHz 4x8GB @ 3333MHz, stock XMP2 settings
GPU: Strix 1080Ti OC
PSU: Corsair AX860i
Chassis: Corsair Obsidian 750D Airflow Edition


My question is that what voltages would you gents recommend for VCCIO, VCCSA, PLL Termination, PCH Core and CPU Standby.

If I leave these on Auto, then IO and SA will go near 1.35v, which seems excessive to me. CPU standby will be near 1.6v on Auto. Should I bring them near to 1.1v, then I still get perfect system stability, measured with RealBench and Prime95. Only odd thing I occasionally get is rather long POST, where memory q-led (yellow one) will remain lit for some time with Q-code 23 (or was it 32?). Sometimes I have to reboot for it to continue, other times it proceeds after dwelling there for about 20 seconds. Normal POST time is 10 seconds, which itself seems a bit long as it is.

As far as I understand, it relates to memory controller and VCCSA voltage, right? So then again, what would be the recommended voltages for aforementioned five variables that seem to overshoot when set on auto and not cause any operating system instability when set really low.


Also, could someone explain this AVX offset to me. Should I leave it on 0, then non-avx workload like P95 v26.6 runs totally great. Should I set any kind of offset and run P95 without AVX, then it BSOD with MACHINE_CHECK_EXCEPTION. Why would it do that if I don't even run AVX load? If I'm not mistaken, then RealBench does utilize AVX, and that will also run perfectly with offset of 0. All hell breaks loose only when I define an offset other than 0.

This post is getting rather long as it is, but bear with me for one more question.

When I do stress testing, then AIDA64, that I only use to graph thermals, not run torture, shows separate core speeds taking a dip downwards, other times it is a lot more unified and flatter graph. Are this just system interrupts coming and going, or is there something off with the system? They don't seems to indicate any instability, just an odd observation. I have attached flatter and jagged screenshots along with my current voltage settings.


If you managed to get this far, than you for reading the post. I hope you could give me some constructive feedback to help me fine tune my system.

Kind regards,
Raimo

[HiVizMan]

First question re voltages.

A. Each CPU is unique. There is sadly no one size fits all. My CPU may need 0.4 more volts to achieve the same stable OC as yours. This journey to find what works best for your system is part of the fun and should be enjoyed. The outcome is often secondary to the journey. Zen overclocking eh? LOL

Second questions (sort of asked it) Memory.
A. The long boot is your system retraining your memory. It may just be that your memory settings are not as stable as you may think. Run a few rounds of memtest and confirm. I like to use RealBench as my stability tester as it does real world stuff and not relying on artificial stressors that actually only tell you that you have passed that test.

Third questions AVX Offset.
A. Again up to you. AVX Instruction Core Ratio Negative Offset: This setting reduces CPU core frequencies by the applied value when an AVX workload is run. The thermal output of AVX workloads is an order of magnitude higher than for non-AVX workloads, which is why this setting has been introduced - so if your system cooling takes a pounding or you are reluctant to have your CPU hammered during AVX heavy applications (do you every do any?) then the option is to reduce the amount of stress on the CPU.

Last question. Your graphs.
A. I can't really see any details sadly. The images are not very high res for me. So going to try and answer based on your text only. Dips could simply be those moments when the stress tests you are doing are handing over from one test phase to another, or they could be related to throttling if your board is not getting enough airflow to cool stuff down. I just don't know.


Last bit from me to you. Mate don't get hung up on those artificial stress tests, they don't actually represent the real world. Do what you bought the PC to do, and if you can do those things without having any errors or BSOD then your system is stable. Let temperature be your guide in all things OC. Current kills, current makes stuff hot - just think a heater yeah.

Enjoy the system.

[Raimo5]

@HiVizMan

Thank you very much for your input on this matter.

Thing is that I don't get the long boot every time, it is very sporadic and might not happen for days at end, or then could happen couple times in a row. RealBench and memtest86 don't find anything to be wrong with memory settings.


That is why I am wondering about VCCIO, VCCSA, PLL Termination, PCH Core and CPU Standby. If I leave them on auto, then they get rather high. If I set them low then I still pass stress tests but I suspect that VCCSA introduces intermittent long POSTs. As mentioned, they are so intermittent that there is no way to change the value and go test it out. I would have to do hundreds of restarts to get a result that would have any meaning, so it's not something I can try out by blindly going after five different voltage values. My Vcore is dialed in and stabile, and I am just wondering what I would have to do to get the POST in order as well.

VCCIO: I don't even clearly understand what it does. Sites say that it is used for all I/O pins except memory related pins. Suggestions say to use 1.15. I have no idea what this voltage is used for, what is considered normal and what are the limits.
VCCSA: Memory controller, PCIe and display engine. I think that could be my issue, so should I try increasing it to, say 1.20v? What is the safe limit for this?
PLL Termination: Again, I don't know what it does, what to set it at and what is safe. Currently at 1.0v
PCH Core: Quite simple, it's voltage for PCH. Currently set to 1.0v and thereabouts when it's on auto as well. Don't really think this is my issue.
CPU standby: Again, not clear on function of this one. Set it to 1.0v as auto shoots it up to 1.6v. Should I increase it?

Regarding AVX, I don't use AVX workloads. I am happy to leave it at 0 offset as long as I don't encounter instability. What puzzles me is why does my system start crashing under stress test when I set it to anything from 1-10 and then run loads that don't even utilize AVX. Is the BIOS broken?


I also have one new question. What is CPU PLLs OC voltage. HWiNFO64 reports it at 2.6v and there is no BIOS setting for it. Is that normal?

I will also upload my graphs to third party picture hosting so you could see them in detail if you are interested. HWiNFO shot from current, idle voltages also included. Click for full size image.

Jagged graph

flat graph

HWiNFO64



Kind regards,
Raimo

[Richo13]

Hi Raimo,

I think your avx offset problem is related to load line calibration & your clock speed.

I have an 8600k/Strix Z370-I & experience the exact same issues as you when using avx offset with certain levels of llc at certain clock speeds.
My system will instantly crash under non avx stress tests if I use llc 6 + avx offset + a core ratio greater than 47 ... although the system is completly stable with avx offset at auto or 0. With the lower level of llc 5 & a core ratio greater than 48 however, the avx offset works exactly as it should, the system is perfectly stable under avx and non avx loads up to 5.1ghz where my chip tops out.

So my suggestion would be to try the next level down of llc and bump your vcore by around 0.020 - 0.030 to compensate for the added vdroop & then try the avx offset again. You may need to go as low as llc 4 even, since you're running at 5.4ghz & the pattern I have come across with my board seems to be the higher the clock speed the lower the llc setting needed for the offset to work.

I dont think this is a bios issue, Im pretty sure its a vrm issue on the board itself. I'm actually pretty surprised that you have this problem with a maximus hero as I thought it may have only been some of the mid to lower end Z370 boards that are effected.

I would be interested to see if that works for you.

[Raimo5]

Hi all,

Thank you very much for your input. Sorry that I haven't been able to reply for a week, but I have been abroad and unable to tune the system.

Now that I'm back, I have set VCCIO to 1.15v and VCCSA to 1.20v, and long POST with rare POST freeze seem to be gone. Well, at least I haven't come across it now and I have ran about 40 power cycles in a row.

I have also managed to get POST time down to two seconds from ten. It took most of this time checking RAM, so I set MRC Fast Boot Auto->Enabled, and it POSTs in two seconds now.

@Richo13, thank you for the interesting tip. I tried it out and indeed it works out for me as well. I set LLC to Level 4 and Vcore to 1.425 to support ~1.288v under load. With AVX offset set to 4, I can easily pass any non AVX workload without crashing and AVX loads as well, running on 5.0GHz respectively. I didn't experiment much, because the system is currently stabile and I am not interested in running that high idle voltages through my CPU.

It's really weird that it makes it work. With Vcore set 1.425 and droops to around 1.300v, LLC 4 allows me to use AVX offset. This voltage is even lower that I normally use under load. I have the system dialed in to 5.4GHz @1.345v Vcore with droop down to 1.314 on LLC Level 7. Why would lower LLC Level enable me to successfully use AVX offset when I'm not even running AVX loads? It just doesn't seem like a VRM issue. Asus even prides themselves that their new VRM designs drastically improve transient load by running twin component power phases, so why would more stabile LLC settings cause this? To me, their new VRM design doesn't really make sense, because putting in only four interleaving power phases, doubling up on components and then calling it twin-eight phase is just lying and is a puny design. I'm not claiming that this is a bad design, because I am running at 5.4GHz, so it can't be that bad, but really, 4 phases on €330 board is just low end and ripoff.

All that considered, and the differences of VRM designs between Z370 and Z390, I can't see how it could be VRM issue. I still think it has something to do with BIOS because lower LLC enabling the successful use of AVX makes no link to VRM as far as I see. After all, LLC 7 is a lot more stabile than 4.

Maybe someone from Asus engineering could chime in on this.

Right now I'm going to keep running 5.4GHz at Vcore 1.345 and AVX offset on 0 with LLC Level 7. I can pass days of P95 26.6, RealBench and Cinebench with their respective, real world AVX loads enabled, and nothing ever goes wrong. Should I ever come to a scenario where heavy AVX workloads impact my stability, then I'm going to revisit this particular issue, but I hope by then Asus have managed to tighten up their problems regarding AVX offset.

Kind regards,
Raimo