[SOLVED] 9700k overheats past 1.210 voltage

[last_intel_guy]

Hello,

Looking to do a 9700K overclock targeting both gaming and game development in Unreal Engine, so I've been using some levels of AVX in my stress test to address the likely use in UE4.

With my OCCT preliminary quick test, I can't run anything above 1.210 volts without getting into 85-95C range on two culprit cores in particular. I've reverted to a 1.210 volt 4.4ghz all core overclock for the time being. Here's my settings and test specs:

Specs:
Intel i7 9700K @ 4.7 ghz (started testing at 5ghz)
Asus z390-P mobo
G.Skill Ripjaws V-Series 32GB (2x16GB) at 3200 Mhz
Noctua NH-D15 (using 1 fan, based on research 2nd isn't completely needed and doesn't fit my case)
Seasonic Focus 750w PSU 80 Plus Platinum

Settings currently for 4.7 ghz attempts:
RAM XMP II Profile (3200 Mhz at 16-18-18-38), no changes besides auto set there
LoadLine Calibration: 6 (since then have tried at 5, still not good enough for higher vcore)
Asus MultiCore enhancement: Disabled (seems to force CPUs in ways manufacturers didn't intend)
SVID Behaviour: Typical
Sync All Cores
AVX Offset: 0
CPU Cache: Auto
Core Voltage: 1.320
VCCIO: 1.1
CPU System Agent Voltage: 1.1
Standby Voltage: Auto
CPU Current Limit: 140% (used to read 170 I swear, may have been updated to max at 140 now)

Settings for 4.4 ghz (stable & cool as of step 5 above testing):
Core voltage: 1.210
4.4 ghz
Load Line Calibration 5 (improved by 3-5 degrees)
All rest are the same.

Here is the usual HWMonitor results from about 5 seconds into the OCCT test (Details of how test is conducted in 1 of above stress tests):
Note: Ignore System Agent voltage result as HWMonitor seems unreliable on this.

I would ideally like to follow the below stress tests as a guide so I am prepared for AVX loads, while OCCT runs hottest with AVX for obvious reasons:

1. OCCT small sets WITH avx (1.250vcore and above high temp results) - 15 to 20 minutes or until temps get too high. (MONITOR CPU ENTIRE TIME) If this passes and does not BSOD or freeze, then continue testing. This works real good for a quick test to see if you should even waste time testing this new setting. If the voltage is not enough you will typically know in under 10 minutes when your comp goes down. If it's OK, then I move on into real testing.
2. Realbench - 1 hour
3. OCCT small sets no avx - 30 to 60 minutes (depending on temps)
4. OCCT large sets no avx - 6 to 8 hours
5. Prime95 small sets no avx - 30 to 60 minutes (depending on temps) (where I have cleared so far currently at 4.4 ghz 1.210v)
6. Prime95 Custom blended with no avx - 6 to 8 hours
7. Aida64 - 6 to 8 hours

Steps attempted:
-Re-apply nh-1 thermal paste to heatsink (seemed to help as previous application was very uneven)
-Update bios -- updated to 2606 from November 2019 as the newest 2808 version seems to have issues according to some OC forums (oh right, yeah I did actually also upgrade to 2808 and then back down to the 11/19 version when it had no advantages)
-Re-arranged case fan layout -- now 3 intake in the front of my 465x iCue RGB, with 1 back exhaust and 1 back top exhaust (used to have a front top exhaust that likely was pulling out the intake), so case balance is +1 intake overall. Probably wasn't necessary but just smart to do.

Any ideas here? all welcome, thanks for reading this far

 

[Lutfij]

Welcome to the forums, newcomer!

Out of curiosity, which version of OCCT are you working with? Perhaps see if reseating the cooler changes your experience with some reviewers stating that mounts/mount pressure does affect temps. Also, what are you ambient room air temps?

Also, I must say, your post is quite thorough for a newcomer to the forums Keep it up!

 

[MeanMachine41]

Hello,

Looking to do a 9700K overclock targeting both gaming and game development in Unreal Engine, so I've been using some levels of AVX in my stress test to address the likely use in UE4.

With my OCCT preliminary quick test, I can't run anything above 1.210 volts without getting into 85-95C range on two culprit cores in particular. I've reverted to a 1.210 volt 4.4ghz all core overclock for the time being. Here's my settings and test specs:

Specs:
Intel i7 9700K @ 4.7 ghz (started testing at 5ghz)
Asus z390-P mobo
G.Skill Ripjaws V-Series 32GB (2x16GB) at 3200 Mhz
Noctua NH-D15 (using 1 fan, based on research 2nd isn't completely needed and doesn't fit my case)
Seasonic Focus 750w PSU 80 Plus Platinum

Settings currently for 4.7 ghz attempts:
RAM XMP II Profile (3200 Mhz at 16-18-18-38), no changes besides auto set there
LoadLine Calibration: 6 (since then have tried at 5, still not good enough for higher vcore)
Asus MultiCore enhancement: Disabled (seems to force CPUs in ways manufacturers didn't intend)
SVID Behaviour: Typical
Sync All Cores
AVX Offset: 0
CPU Cache: Auto
Core Voltage: 1.320
VCCIO: 1.1
CPU System Agent Voltage: 1.1
Standby Voltage: Auto
CPU Current Limit: 140% (used to read 170 I swear, may have been updated to max at 140 now)

Settings for 4.4 ghz (stable & cool as of step 5 above testing):
Core voltage: 1.210
4.4 ghz
Load Line Calibration 5 (improved by 3-5 degrees)
All rest are the same.

Here is the usual HWMonitor results from about 5 seconds into the OCCT test (Details of how test is conducted in 1 of above stress tests):
Note: Ignore System Agent voltage result as HWMonitor seems unreliable on this.

I would ideally like to follow the below stress tests as a guide so I am prepared for AVX loads, while OCCT runs hottest with AVX for obvious reasons:

1. OCCT small sets WITH avx (1.250vcore and above high temp results) - 15 to 20 minutes or until temps get too high. (MONITOR CPU ENTIRE TIME) If this passes and does not BSOD or freeze, then continue testing. This works real good for a quick test to see if you should even waste time testing this new setting. If the voltage is not enough you will typically know in under 10 minutes when your <Mod Edit> goes down. If it's OK, then I move on into real testing.
2. Realbench - 1 hour
3. OCCT small sets no avx - 30 to 60 minutes (depending on temps)
4. OCCT large sets no avx - 6 to 8 hours
5. Prime95 small sets no avx - 30 to 60 minutes (depending on temps) (where I have cleared so far currently at 4.4 ghz 1.210v)
6. Prime95 Custom blended with no avx - 6 to 8 hours
7. Aida64 - 6 to 8 hours

Steps attempted:
-Re-apply nh-1 thermal paste to heatsink (seemed to help as previous application was very uneven)
-Update bios -- updated to 2606 from November 2019 as the newest 2808 version seems to have issues according to some OC forums (oh right, yeah I did actually also upgrade to 2808 and then back down to the 11/19 version when it had no advantages)
-Re-arranged case fan layout -- now 3 intake in the front of my 465x iCue RGB, with 1 back exhaust and 1 back top exhaust (used to have a front top exhaust that likely was pulling out the intake), so case balance is +1 intake overall. Probably wasn't necessary but just smart to do.

Any ideas here? all welcome, thanks for reading this far

Your cooling is inadequate for a 5.00Ghz Overclock. A H110i AIO or equive would fix that if it will fit your case?

I would apply a negative offset of -2 voltage on the two cores that are overheating. Core boost can raise temps on those especially when using AVX apps

LLC should be set to 7 for 5.00GHz to prevent droop.
And there's more.

 

[last_intel_guy]

Welcome to the forums, newcomer!

Out of curiosity, which version of OCCT are you working with? Perhaps see if reseating the cooler changes your experience with some reviewers stating that mounts/mount pressure does affect temps. Also, what are you ambient room air temps?

Also, I must say, your post is quite thorough for a newcomer to the forums Keep it up!

Thanks! OCCT version is 7.0.4, latest I think. Running that and HWinfo64 as admin, all other programs closed. Not messing with program priorities atm though in Task Manager.

Actually -- the first thing I did was reseat the cooler. I really botched the thermal paste first time it looked like (luckily only limited stress testing before that), and seemed to be much better with an application following Noctua's guide so that should be good now.

Now that you mention it.... the heatsink did seem tight but I will crank in the screws just in case I was a little cautious on too much pressure. Or do you mean the 4 screws that tighten the mount for the heatsink? Assume I would have to redo thermal paste again to take off sink and reach those screws.. but I am almost positive I checked that when I reapplied paste.

Keeping room temps around the 70-72 ambient mark. no humidity either (west coast).

Your cooling is inadequate for a 5.00Ghz Overclock. A H110i AIO or equive would fix that if it will fit your case?

Thanks for the tip! From my research the Noctua DH-15 (best air fan you can buy) has actually outperformed water coolers and is specifically designed for full 9700K overclock capability, I think maybe even outperforms that one you've mentioned, saw on LTT. Did you say that just because it's air cooling or specifically this model of fan?

LLC should be set to 7 for 5.00GHz to prevent droop.
And there's more.

Also LLC has been a struggle - I actually isolated this variable and ran tests at level 6 as well and the temps seemed much worse at higher values. I believe this may be since my cheap asus mobo ($150) does not sense and set voltages/cpu temps very accurately. Since LLC adds its own resistance calibration to mitigate the inaccurate sensors, lower (5) is better in my case it seems where it actually supplies voltage at the correct level due to vdroop compensation for the amount that the sensor is off. I've gone back and tested up to LLC 6 several times but still overheating so far. Buildzoid had a great vid to explain this. I don't think my mobo can run at 7 safely.

I would apply a negative offset of -2 voltage on the two cores that are overheating. Core boost can raise temps on those especially when using AVX apps

Regarding the voltage offset -- you may be right but I do not think my cheap mobo allows such fine tuning, do you just mean offset the individual core frequency of those 2 cores at -2 less? Funny thing is sometimes cheap mobos can just be reading the temp sensors wrong from what I hear. Is there other settings I would need to change to run the CPUs lower?

I should also update: I actually have now been able to get up to 1.305 volts for a 4.7 ghz overclock. Up to stress test 5 on 4.7 ghz now (Temps max at 90 with OCCT AVX testing for 15 min).

I found the culprit to be leaving Standby Voltage and VCCIO to auto, changing those to 1.0v seems to cool the temps considerably. STILL can't get above 1.305 though for AVX tests! Note: Tried a 4.9 ghz test with -2 AVX offset and was quickly not stable.

I think this means to get cooler I would need to set manual voltages for these Ai Tweaker settings most other guides say to skip (and also skip VCCIO and Standby), but I am not very familiar with standard i7 voltages here. Just seems that my Asus z390-P is too cheap to manage its own voltage where other manual voltage overclocks can even leave it to auto:

Realtime Memory Timing [Disabled]
FCLK Frequency for Early Power On [Auto]
VPPDDR Voltage [Auto]
Internal PLL Voltage [Auto]
GT PLL Voltage [Auto]
Ring PLL Voltage [Auto]
System Agent PLL Voltage [Auto]
Memory Controller PLL Voltage [Auto]
CPU Core/Cache Current Limit Max. [255.50]
Ring Down Bin [Auto]
Min. CPU Cache Ratio [Auto]
Max CPU Cache Ratio [Auto]

Any help on testing some of these above voltages with manual values, or feedback to my above? Appreciate thus far!

 

[last_intel_guy]

Latest update to revise the above, thanks everyone for the comments which have been super helpful:

Several signs point to a mismatch in RAM with my motherboard, using a pretty cheap Asus z390-P and trying to run my DDR4 dual channel RAM at the manufacturer's XMP profile 3200 Mhz (XMP II, not XMP I)

The good news is tempwise I now am able to reach 1.365v core voltage without overheating on temperature tests at full load with AVX on OCCT (small sets). I can keep temps in the 85-87 degree range on my 2 hottest cores (I know it's high, but much lower with non-AVX small sample tests, never over 80 as all guides note) with the following adjustments since my last post. Note that I made each of these changes individually and did not do anything in combination as to ensure clear understanding of what was affecting each test (repeating Step 1 with HWinfo64 monitoring up to the 4.9 ghz threshold):

Updates/Changes:
-LLC = 4 (drastically improved temps, LLC 5 won't work at all past 1.300v)
Intel TVB = Disabled (no major impact here but according to online it seems to affect voltage fluctuation)
-Re-mounted and re-pasted nh-1 to heatsink (noticed that screwing in heatsink this time felt more even, the mount backplate may have been off before). This alone did not seem to cause a temperature change, however, so may have already been optimal.
-Did a small fan adjustment on CPU noctua fan to blow into the VRM (by moving it lower to blow between heatsink and motherboard). I noticed on the z390-P board the VRM has some kind of metal cover, I assume this is a helpful heatsink so haven't touched it besides adding that airflow.
-AVX Offset: 3 (actually not seeming to make much of an impact on temps, maybe 1-2 degrees max)
-Upped to 4.9 ghz after several Step 1 tests showed no stability or temperature issues, which made it all the way to Step 5 before Prime95 found BSOD crashes.

Should my CPU cache be running auto at 4300 Mhz, or is that a bad sign?

Also seems there is some sort of voltage offset possibly from the Loadline Calibration, booting into BIOS usually shows 1.350 vcore instead of 1.365.

With XMP II on, my DRAM voltage is at 1.350 volts which shows as "yellow" in the BIOS, which I know is a warning that it's high. Would this be affecting stability/voltage for my CPU, and if so should I just run the RAM at stock? For game engine work I like high speed RAM, so I don't know much about that tradeoff from 2133 Mhz to 3200 Mhz but assume it may be a lot.

I am seeing at times in HWinfo as well that core frequency is fluctuating from 4900 to 4600 Mhz on non-AVX tests, and am using the Sync All Cores setting. May try Per Core setting but not sure if that will negatively impact performance for 8-core processing in full loads on Unreal Engine down the road.

Further DRAM evidence is shown as I BSOD when I hit the Prime95 30-60 min small FTT tests, usually 10-15 mins in with a crash dump file, temps stay under 80. I think I even saw MEMORY as a reading in one. Passed all tests leading up to that on the changes above, even OCCT 6-8 hour overnight. Seeing kernel32 so seems like voltage could be involved.

Currently running Memtest86 through at least 2 full passes with both sticks in to see if an error is triggered at XMP II (with CPU reverted to stock turbo settings, don't think that matters though), if not then I'm not sure what the issue may be or what my next test tweak would be.

 

[last_intel_guy]

After tweaks it seems like the key answer to my initial issue was the voltage for the System Agent and VCCIO. With all other variables tested under the initial issue, this seems what improved temps as well as adding the standard other voltages previously on Auto as noted by Intel (PCH Core, Standby voltage).

Going to choose the re-seating the heatsink as the Best Answer since I have suspicions that while I did not greatly improve any temps by doing this, some weird screw tension on my heatsink mount before was nearly nonexistent after, and something not a lot of people talk about in mounting the Noctua DH-15. I wouldn't be surprised if the pressure was uneven across the cpu, maybe only slighty but I'm sure it helps. Recommend anyone using this fan actually stand the motherboard up in the case (if the case has an opening on the backside) and screw on the mount with it vertical, so there is no gravity to make the mount screw in uneven.

Culprit here is likely that I've also not found a good AVX stress test (doing small set AVX on OCCT is a mixed bag according to more experienced community members). Something (CPU or 3200 MHZ XMP RAM) is not getting enough voltage and lots of crashes.

Not quite satisfied with my overclock yet but temps aren't the issue now with my LLC at 4 and AVX offset at -2 of a 4.9 ghz overclock. Guess I'll buy a better mobo next time than the cheap Asus option.

For an 8 hour Realbench test at max 75C core temp, we can call this main question solved