Problems with reaching steady and constant CPU package power

[Dueholm]

Hi guys!

I’m in the process of benchmarking thermal performance for several different cooling solutions and I’m facing troubles with setting up the BIOS properly. For scientific reasons, I like to obtain a constant CPU package power throughout a period of 30-60 minutes. Also, I’d like to reach a mild overclock of about 4.7 GHz (the exact value is not important). The CPU Package power must not reach above 170 W, as small coolers should be able to adequately cool the CPU. For me, it’s important to be able to baseline thermal performance for multiple cooling solutions, using the same BIOS settings.

With my current BIOS settings (I’ve listed settings below) I’m not able to reach a steady CPU Package power. It seems to be somewhat constant for the first 15 minutes, then increases for the next 15 minutes, followed by a decrease and so on. I’ve attached a figure of the CPU power as a function of time.

Can anyone explain this behavior to me and an idea on how to reach a steady and constant power (if it’s even possible!)?
I’ve tried messing around with a few BIOS settings, such as the LLC, but have found no success.


Thanks in advance!

View attachment 212103

-----------------------------------------------------

Additional information:
Stress testing software: OCCT Enterprise v6.1.0 @Small data set, 8 threads, AVX2
Monitoring Software: HWiNFO64

Hardware specs:
CPU: intel i7-9700K
Motherboard: ASUS ROG STRIX Z390-E GAMING
RAM: Corsair VENGEANCE LPX DDR4 - 16GB (2x8GB)

BIOS settings:
BIOS version: 1602 x64
AI Overclock Tuner: XMP II
BCLK Frequency: 100.0000
ASUS MultiCore Enhancement: Disabled
SVID Behavior: Typical Scenario
AVX Instruction Core Ratio Negative Offset: 0
CPU Core Ratio: Sync All Cores
Core Ratio Limit: 47

CPU Load-line Calibration: Level 4
CPU Current Capability: 140%

Intel® SpeedStep™: Disabled
Long Duration Package Power Limit: 4095
Package Power Time Window: 127
Short Duration Package Power Limit: 4095

CPU Core/Cache Current Limit Max.: 255.75

CPU Core/Cache Voltage: Manual Mode
CPU Core Voltage Override: 1.250
CPU VCCIO Voltage: 1.10000
CPU System Agent Voltage: 1.10000

 

[EarthDog]

You would want to manually set clock speeds and voltage. Manually set the CPU to a specific clock speed, and manually set the voltage. If you've done that, I would look at a different stress test, perhaps. I don't run OCCT, but I know running AIDA64, the results are as consistent as they can be (within several watts) over 30 minutes (how long we test). There will always be some kind of fluctuation though.

In AIDA64 once it's ramped up to a steady-state, there is still several W of fluctuation. That's about as good as it is going to get. You may want to consider logging the power use and using an average once at a steady-state to determine the wattage value.

 

[Dueholm]

Thanks, EarthDog! I'll try using AIDA64, to see if it's a software-related issue.
I can cope with consistent fluctuations by averaging the power and other desired properties, but significant drops or increases in CPU power dissipation will in some cases lead to systems that'll not reach a steady state, especially systems containing large amount of water.

I've manually set the clock speed to 4.7GHz with a 1.25V.

 

[EarthDog]

Are these several watts going to matter in the end? Offhand, I'm not sure what would be any more stable. Some stress tests do different things at different times (which may explain the results in OCCT)... but I know for a 30 min test in AIDA64 (with CPU/FPU/Cache checked), power draw stayed within about 6-9W

Looking at your chart more closely (the axis on the left), once the CPU ramps up your +/- is only 3W either way. At worst your variance is <3%. Best...is <1.5%. You won't do much better than that without a hot plate or something where you can control the load more precisely. These few watts won't make or break saturation. It's likely your measuring tools aren't accurate enough to even see that small of a change temperature wise.

The more liquid you have, the longer it will take to reach that equilibrium though.

That chart is odd.... why does it take several minutes for your CPU to ramp up? It's typically instant/a second two to see it....

 

[Dueholm]

Thanks again, EarthDog!
Having a constant and consistent power matters a lot to me in order to baseline thermal performance equally for all cooling solutions. I just did some quick calculations on it's importance and found a significant difference, depending on where you're averaging your measurements. As you can see from the below attachement, the thermal performance differ by almost 2*C, even though I'm normalizing the CPU Package temperature to account for differences in CPU power draw and ambient temperature.

View attachment 212124

I've now conducted tests on several platforms (AM4 3950X, intel i7-9700K and intel i9-9900K) all showing similar behavior (increase and decrease in CPU power).
I've looked for plausible reasons for this behavior and have come down to the conclusion that it's a software related "issue". As the majority of stress testing software are designed as stability checkers, it'll run different operands.
For OCCT I've recieved confirmation that the operand switched every 15 minutes. I believe Prime95 and Aida64 to do it similarly.
Do you happen to know of any software that does not switch operand?

 

[EarthDog]

Hmmm....dont think aida64 switches. My wattage BOUNCES around for the 30+ mins we test. Your chart shows distinct differences in chunks, aida doesnt seem to. If I chart my power use, it just fluctuates +/- a few watts either way.

As I said, if you want a constant load, you may need to make a hotplate.

Your table... are those the same cpu? Is your cpu trmp listed peak or average? Did ambient stay there throughout each test? In my office, i have a window ac unit otherwise my temp changes a couple of C in long stress tests like that.

Also, isnt thermal resistance a constant? In that so long as the properties of the cooler and enviro factors don't change, that should stay the same, right?

EDIT: Zooming in on my data, It looks like its all over the place. But it could be the polling frequency (1s)? Not sure...

View attachment 212125

 

[Dueholm]

Yeah interesting... it seems like your system does not experience the same behavior. May I ask what kind of settings your using i Aida64? Below is my measurements during Aida64 testing, where "chunks" are still present.

View attachment 212126

I've tried using hot plates, but doing so results in different issues, as a cooler perfoming good on a hot plate does not necessarily perform as good on a CPU.

The data in my table is for the same CPU running the same stress test, though the properties are determined at different times (time periods corresponding to the power level "chunks"). They are all averages.
Ambient temperature was fairly constant, as all data are from the same test.

The thermal resistance is calculated as: R = (T_CPU - T_Ambient)/P_CPU.
Its only constant as long as the CPU operates in an identical way, but as operands changes so does the CPU behavior and therfore also the thermal resistance. Though, if testing on a hot plate, you're correct, and thermal resistance would be constant.

Polling frequency is 2s.

 

[EarthDog]

AIDA64 Stress test settings = CPU/FPU/Cache checked (see post4).

Are you testing the cooler or the CPU? A hot plate will give set of generic results that Cooler A is better/worse than cooler B at xxx loads. When you put a CPU underneath, you add variables to it. Each CPU is different as far as its own temps anyway. Silicon quality, TIM application between die and iHS (as well as your application on top of the IHS. 2C difference is just outside of margin of error really.

But a heatsink's thermal performance will not change with meager load differences like that. It's still removing heat at the same pace.

 

[Dueholm]

Thanks, though there were more settings to configure.

I'm testing the cooler. Using a hot plate does not necessarily tell which cooler is performing the best when actually installed on a CPU. Most coolers are optimized for a specific processor, e.g. i7-9700K, thus optimized to provide the best cooling for the specific die locations. As a hot plate cannot accurately account for die locations, this optimization is not included in the result. But I agree that testing on a cold plate will yield an indication of which cooler is the better.

As I use the reported die temperature to calculate the thermal resistance, I’m not only finding the thermal resistance of the heat sink but a combination of the CPU, interface material, and heat sink resistance (in the figure that’s ψ_jc + ψ_cs + ψ_sa). As die behavior changes so does the combined thermal resistance. It’s true, that the thermal resistance of the heat sink remains the same.
If I were to find the heat sink thermal resistance (ψ_sa), I needed a thermocouple at the bottom of the heat sink. It’s possible to do, though difficult, the problem is that modifying the heat sink might change the thermal performance.

View attachment 212128

 

[EarthDog]

Most coolers are optimized for a specific processor, e.g. i7-9700K, thus optimized to provide the best cooling for the specific die locations.

Are they? That's news to me. I've seen updates to larger cold plates because CPU size (HEDT) has increased, but I don't think there are specific coolers optimized for specific processors. Most (all?) out there fit a slew of platforms, AM4/TR2, Socket 115x,1200,2066, etc... they all come with brackets to fit everything. Some water cooling blocks had 'jet plates'

IMO, it's better to use a hot plate. As I said and you've shown, there are multiple variables involved with a CPU. You're testing heatsink performance with a given load. That 'relative' performance carries over regardless of the variables that change underneath. Everyone will get a different result because of all those variables, but the better heatsink is the better heatsink.

Anyway, AIDA, for me, seems to be as steady-state as you're going to get. I can't think of anything offhand that would run as long as you need it with a constant load. You can try other stress tests.. Intel XTU has one. Linpack...all solid.... just not sure of its loads.

I'm really interested to see what you come up with.

 

[Ben333]

OP you may find these interesting: (in regards to using a "hot plate" aka constant load aka die simulator)

https://www.overclockers.com/heatsink-testing-methodology/

http://www.procooling.com/index.php?func=articles&disp=140

https://www.frostytech.com/articles/1915/index.html

 

[Dueholm]

Hi again. Sorry for the late response, I've been working to further understand and solve this issue. It has actually been a quiet an interesting process

I've figures that the reason for variations in CPU Package Power were due to the switching of operands. I've been in contact with the guy making OCCT, and the newest OCCT version now features a setting called "Steady load", which does not switch operands. This setting is perfect for consistently determining thermal performance for any cooling system.

Thanks for the great help guys, which let to this conclusion!

 

[nschilperoort]

Hello Dueholm,

I bumped into your thread and noticed you are running the same chip and very similar motherboard...I've got the 9700k running in an Asus Prime Z390-A. Having recently completed this build I wasn't too familiar with the default BIOS settings for overclocking, and as I'm sure you know they're 100% crap. Found my cores spiking to high 80s and mid 90s playing Horizon Dawn on Ultimate Settings, etc.

That was about a week ago and since then I've been searching for OC settings that result in a computationally and thermally reliable setup. Earlier today I tried the settings you posted in your first message on this thread and found that for most stress tests and benchmarks they worked 100% great! No more unexpected blue screens, no more run away core temps and very reasonable voltages. Not sure how you came up with those settings (curious to hear), huge thanks for posting!

That said, the only exception is Prime95 running the blended tests with AVX instruction sets. After attempting to run Prime95 I was forced to modify the AVX Instruction Core Ratio Negative Offset to 2 instead of 0. This allowed me to run Prime95 for 30 minutes with no errors (ran for 30 min after which I forced quit) however thermals were still peaking to 94 on one of the cores. So I dropped the Core Ratio Limit to 46 and and increased the AVX Instruction Core Ratio Negative Offset to 3 and now I'm running Prime95 with temps no higher then 90 on one of the cores and the remaining cores are all in the mid to high 80s. These temps are only spiking occasionally on the tests with shorter string lengths < 10k.

Anyways, I suspect my CPU cooler is a little under spec'd...it's the be quiet! Shadow Rock 3 air cooler. I'm keen to try something more industrial like something from the Be Quiet Dark Rock series or the Noctua NH-D15. My hope would be to match your BIOS settings and still see reasonable core temps and no errors on Prime95. Couple questions for you...

1. What cooler are you running ?
2. What core temps are you seeing ?
3. What CPU stress tests are you using ?

Thanks again!

 

[nschilperoort]

A day later a day wiser. I've since dropped Vcore to 1.2V and reduced AVX Instruction Core Ratio Negative Offset from 3 to 2. End result is all 8 cores are pinned to 4600 MHz for regular work loads and 4400 MHz for AVX instruction sets. Temps now are much better when running Prime95 with same fan profile and same CPU cooler as before. The hottest core doesn't go above 84 degrees C for most stressful torture test Prime95 has to offer after running it for about an hour. Technically I should let it run overnight for more rigorous test however at this point I'm just happy I've stumbled on what seems to be a computationally and thermally stable OC'd Bios setup.
Combined with a RTX 3080 GPU my 3DMark scores are a hair shy of 15000, not bad for a OC newbie I'd say. That said I'm really not interested in breaking any records, more focused on finding a fast, stable and reliable setup.

BTW...I tried bumping up the Core Ratio Limit to 47 however after running Cinebench for about a minute or so the computer froze. Tells me I'm short on voltage...that however leads to hotter cores and new CPU cooler which I'm really not too keen at the moment. Maybe later this year, we'll see.