Question 9700k is too hot! Or is it?

cholzman

Prominent
Aug 25, 2018
8
2
515
I have a 9700k (on a Gigabyte Z390 Gaming X motherboard with latest BIOS as of Feb 19, 2019) hitting 74c within 4 minutes of Prime 95 v29.4 Build 8 x64. I like to let it run with default values for 15 minutes to ensure it never breaks 85c. But it hits 95c soon after and I stop the test before the 15 minutes completes.

So, I read these posts and I grab Prime 95 v26.6 Build 3 x64 and, using default settings, it runs the full 15 minutes with a max CPU temp of 67c.

In my case, disabling AVX as instructed on Prime 95 v29.4 Build 8 x64 resulted in no difference. I was still unable to go the full 15 minutes.

Then I noticed something - the default test settings are different from version 26.6 to version 29.4.

26.6 Default Test Settings
prme-95-266-default-settings.jpg


Notice the shaded out area - Memory to use (in MB)=1600, and Time to run each FFT size (in minutes)=15

29.4 Build 8 Default Test Settings
prme-95-294-default-settings.jpg


Notice the shaded out area - Memory to use (in MB)=2000, and Time to run each FFT size (in minutes)=3

By choosing CUSTOM and reversing each version with the others default settings, the temps are adversely affected.

Turning off AVX made very little difference.

Turning off all options, Prime95 shuts itself down (latest version). Turning off all options except CPUsupportsSSE2=0 then allows the program to run without shutting itself down, but even still, no difference in temps. Only by changing the 29.4 default test settings to match 26.6 default test settings am I able to run the program for the full 15 minutes in a safe temperature zone (74c max in this case). Changing the 26.6 settings to 29.4 defaults result in much, much higher temps.

Can anyone explain why the default test settings were changed and why disabling AVX makes very little difference? (again, disabling AVX only on the latest version of Prime 95, just to be clear)

Thank you!
 
Last edited:
Not sure why the later versions are a lot more 'torturous', but, only Blender, Cinebench, or similar rendering loads, etc., would normally ever generate such a CPU load as heavy as even what is found in V26.6...

Perhaps that 'FFT time', being drastically shortened in v29.4, is indeed much more akin to a 'full throttle' processing load....

My 7700K easily showed 10-12C differences in peak temps achieved with assorted P95 versions (v26.6 vs. V29.4), even with a Noctua NH-D15....(I hit 86C with V29.4 in less than 5 min)

ANyway, as a point of comparison, your CPU got better temps than mine in v26.6....

So until we get definitive answers on what is an 'unrealistic full throttle load' wtih P95 versions, I will stick with 26.6 blended, thinking mid-60's in that version is decent....
 
  • Like
Reactions: cholzman

cholzman

Prominent
Aug 25, 2018
8
2
515
Isn't it odd that the video I made to demonstrate the test is removed by the moderator as being 'self-promotional' in my original post starting this thread?

Also, I've authored articles for Toms Hardware in the past, so its odd that the moderator of this forum assumed so much incorrectly, didn't verify and then acted in such a way?

For anyone interested, just search for the title of the video, Benchmark Axioms, on YouTube. I'm reporting the moderator.
 
Last edited:

CompuTronix

Intel Master
Moderator
cholzman,

AVX can be disabled in Prime95 versions later than 26.6 by inserting "CpuSupportsAVX=0" (without quotations) into the "local.txt" file in Prime95's folder.

If done correctly, AVX versions 27.7 through 29.4 will have the same thermal behavior as non-AVX version 26.6. It really does work.

Exactly how did you disable AVX, and what did you modify to do it ... character per character, please?

Utilities that don't overload or underload your processor will give you a valid thermal baseline. Here’s a comparison of utilities grouped as thermal and stability tests according to % of TDP, averaged across six processor Generations at stock settings rounded to the nearest 5%:



All tests will show 100% CPU Utilization in Windows Task Manager, which indicates processor resource activity, not % TDP workload. Core temperatures respond directly to Power dissipation (Watts), which is driven by workload. Prime95 v26.6 Small FFT’s provides a true and steady 100% workload, so if Core temperatures are below 85°C, then your processor should run the most demanding real-world workloads without overheating.

Are you using AVX offset in BIOS? Do you use any AVX apps? Do you need to be AVX stable? If not, then accept the results from 26.6 and enjoy your rig.

As per Intel’s Datasheets, TDP and Thermal Specifications are validated “without AVX”.

You might want to read this: Intel Temperature Guide - http://www.tomshardware.com/forum/id-1800828/intel-temperature-guide.html

CT :sol:
 

cholzman

Prominent
Aug 25, 2018
8
2
515
Yes, I saw your post on a different thread and you seem to be the most knowledgeable on this topic. I did disable AVX in the latest version of Prime95, as I explained in my previous post, and the CPU still exceeded 85C under 7 minutes. However, when I changed the default test settings to match the default settings in 26.6, then Prime95 ran the full 15 minutes and never exceeded 85C. Again, this is a 9700k. My video details all of this and I encourage you test it yourself.

(and for anyone following along, there is no local.txt file until you run Prime95 for the first time.) I would appreciate if you could include that information as I was baffled why I had no local.txt file.

Thank you!
 

CompuTronix

Intel Master
Moderator
I did disable AVX in the latest version of Prime95 ... My video details all of this and I encourage you test it yourself ...
I viewed your video, which my colleague wouldn't have deleted had there not been a commercial lead-in. During viewing, I found a few inconsistencies which I'd like to point out.

In addition to testing every non-proprietary stress utility available since 2006, I've also tested different versions of Prime95. Since the release of the first AVX version 27.7 in May, 2012, "CpuSupportsAVX=0" really does work. Here's the version history - https://en.wikipedia.org/wiki/Prime95

Prime95's default test "Blend" is a fluctuating workload best suited for testing memory stability. Although you highlighted the differences between P95 versions in their default settings for "memory to use" and "time to run", the amount of memory used in either version of Blend is too large to fit within the CPU's L2 cache, which makes Blend primarily a memory intensive workload. Accordingly, time to run affects the intervals at which Blend will cycle between varying FFT sizes, which are shown in the text lines for each thread.

Larger FFT's that don't fit into cache are memory intensive workload which have lower Core temperatures. Smaller FFT's that do fit into cache are CPU intensive workloads that drive the highest Core temperatures. Different P95 versions of Blend will cycle through varying FFT sizes at different times, (for example 3 minutes versus 15 minutes), which causes workloads and Core temperatures to fluctuate.

These differences in thermal behaviors become clear when visualized by observing Prime95 workloads on a graph, rather than focusing on peak numerical thermal values.

The "Charts" in SpeedFan span 13 minutes, and show how each test creates distinct thermal signatures.



Shown above from left to right: P95 v26.6 Small FFT's and Blend, followed by Linpack and IntelBurn Test.

Note the steady thermal signature of Small FFT's, which allows accurate measurements of Core temperatures. A steady 100% workload is key for thermal testing so the CPU, cooler, socket, motherboard and voltage regulators can thermally stabilize.

SpeedFan has colorful graphs that are easy to visualize, however, the numerical values may require an offset correction of 15°C, which is found under Configure > Advanced > Chip > Intel Core. Once individual Cores are configured to a known accurate utility, Core temperatures in SpeedFan will be accurate. Hardware Info is highly detailed, frequently updated, known to be very accurate, and is instead preferred over Hardware Monitor. A drop down menu will appear in Hardware Info that features "Show Graph" when you right-click on any temperature parameter such as "CPU Package". AIDA64 and OCCT also have graphs. Although their stress tests are OK for "stability" testing, I don't recommend using them for "thermal " testing, as their workloads are inappropriate for that purpose.

SpeedFan - http://www.almico.com/sfdownload.php
Hardware Info - https://www.hwinfo.com/download.php

When running Hardware Info, use "Sensors-only".

For testing thermal performance, it's instead recommended that you use only P95 v26.6 Small FFT's, which is a proper steady-state 100% TDP workload that yields steady Core temperatures, as shown in the image above. Since the amount of memory used for Small FFT's fits in the CPU's cache, Core temperatures are negligibly affected by differences in time to run Small FFT's of various sizes.

Except for potential differences in CPU microcode and BIOS settings such as Core voltage (Auto, Adaptive, Offset, Manual), I'm uncertain as to why CpuSupportsAVX=0 apparently isn't working for your 9700K in 29.4's Blend test. Respectfully, I suggest that you scrutinize the character string, character per character, as simple mistakes can happen to the best of us.

I would look for something like:

CpuSupportAVX=0 (no "s" after "Support"; should be "Supports")
or
Cpu Supports AVX =0 (space between "X" and "="; should be "X=")

If your character string is indeed correct, then try 29.4 using only the Small FFT's test.

I would appreciate if you could include that information as I was baffled why I had no local.txt file.
I have edited your suggestion into the Temp Guide ... thank you!

The Intel Temperature Guide explains how to properly set up and run a CPU thermal test using Prime95 v26.6 Small FFT's, which replicates a 100% TDP workload in accordance with Intel's Datasheets. See Sections 10, 11 & 12.

Again, as per Intel’s Datasheets, TDP and Thermal Specifications are validated “without AVX”.

If you use AVX apps such as for rendering or transcoding, those real-world workloads may seldom reach that of pre-AVX Prime95 v26.6 Small FFT's. Later versions with AVX such as 27.7 through 29.4 have completely unrealistic workloads, and thus Core temperatures, just as you emphasized in your video.

CT 😎
 
Last edited:

cholzman

Prominent
Aug 25, 2018
8
2
515
Wow! This is super-fantastic information! Thank you for taking the time to explain all of this. I can assure you, with regards to disabling AVX as instructed, I tried numerous times and different techniques just to be sure if capitalization mattered, etc... It did make some difference in temps, but the CPU still exceeded 85c well under 15 minutes. It is not my goal to complicate my life, but rather to ensure some level of quality control that increases the statistical likelihood that the PCs I build for my clients will withstand whatever the client throws at it. I don't want the QC to take too much time (customers want everything yesterday) but I also don't want to rush the PC out the door, only to have it come back and take more of my time and have an upset and inconvenienced customer. To that end, I now run Memtest86 through one full pass to ensure all the RAM is verified as functional and then I run Prime95 for 15 minutes to ensure the CPU never exceeds 85c.

I tried running Prime95 26.6 with small FFTs, but the CPU hit 87c at 12 minutes, so I stopped the test.

I tried running Prime95 29.4 Build 8 with small FFTs, but the CPU hit 100c in 45 seconds, so I stopped the test.

I tried running Prime95 29.4 with small FFTs and CpuSupportsAVX=0 in the local.txt, and the CPU passed the 15 minute test with a max temp of 82c

I tried running Prime95 26.6 with default BLEND and after 15 minutes, CPU hit 71c max.

I tried running Prime95 29.4 Build 8 default BLEND but the CPU hit 100c in just under 5 minutes.

I tried running Prime95 29.4 Build 8 with default BLEND and CpuSupportsAVX=0 and, after 15 minutes the CPU hit 85c max.

So, I stand corrected that AVX did make a difference, but it was not equivalent to the default BLEND test in 26.6, which is how I interpreted a previous post. In reality, disabling AVX resulted in cooler temps from 29.4 than 26.6 in the isolated small FFTs test, as the results show above.

Perhaps the miscommunication is in the precision of NOT using default test settings. I wrongly assumed that the instructions to run FFTs was all that was important, and by selecting other tests at the same time (BLEND) would make no difference (why would testing the RAM sub-system affect the CPU sub-system? That was my logic at the time. I see it differently now.)

Do you have any opinion or suggestion with regards to what and how I QC new builds and/or any suggestions to optimize the process?

(as for the advertisement that runs before the video, that is YouTube, not me. If I charged for the video, then I could pay for the bandwidth to distribute it independently. Heck if the forum editing tool allowed me to upload a video here, I would, but clearly they don't want to pay for the bandwidth, either. So how exactly does one distribute a video designed to educate and communicate faster, and with more clarity, in this forum?


Thank you!
Carey
 
Last edited:
Mr. H,

I wonder what results the above 9700K would show when/if fitted with a premium air cooler? Perhaps your sample busting above 85C w/small FFTs within 12 min might indicate only that that the chosen cooler does not quite cope with the thermal load as well as some other air coolers, and, although that cooler might perhaps perhaps have been adequate in the past for a 7700K, 9600K, etc, said cooler now simply might not 'cut it' with a 9700K (or above models, obviously), despite various CPUs allegedly having similar 'TDP' values. (clearly, a 9900K cranks out more MUCH more heat than an 7600K or 8600K) So before we decide perhaps P95 version a/b/c is perhaps too stressful, perhaps we can find results of your CPU using other air coolers...(stock, and with MCE-enabled)

Small FFTs enabled only seem to add about 4-5C to the peak temps w/ my 7700K, raising temps to about 76C peak, normally sitting w/ various cores and package temps at 73-74C (MCE enabled, all core 4500 MHz turbo operation during P95 v26.6)

(I thought I saw all- core operation in your video at about 4.6-4.7 GHz as noted in HWMonitor? Imagine what will happen to temps if someone is using an MCE-capable mainboard, which will shoot for all-core turbo clocks at 4.9 GHz...; I'd imagine most 'enthusiasts' will do exactly that, with actual power dissipated likely now routinely at more the 135-150 watt peak levels.)
 
Last edited:
  • Like
Reactions: cholzman

CompuTronix

Intel Master
Moderator
... It is not my goal to complicate my life, but rather to ensure some level of quality control that increases the statistical likelihood that the PCs I build for my clients will withstand whatever the client throws at it ... I run Prime95 for 15 minutes to ensure the CPU never exceeds 85c. ...

I tried running Prime95 26.6 with small FFTs, but the CPU hit 87c at 12 minutes, so I stopped the test. ...

I tried running Prime95 29.4 with small FFTs and CpuSupportsAVX=0 in the local.txt, and the CPU passed the 15 minute test with a max temp of 82c. ...

So, I stand corrected that AVX did make a difference, but it was not equivalent to the default BLEND test in 26.6, which is how I interpreted a previous post. In reality, disabling AVX resulted in cooler temps from 29.4 than 26.6 in the isolated small FFTs test, as the results show above. ...

That's quite puzzling. At a steady ambient temperature, I've never seen more than 2°C variation in default Small FFT sizes between P95 versions throughout a 30 minute run. Differences in lower order variables such as CPU microcode, and stock versus custom BIOS settings for Core voltage (Auto, Adaptive, Offset, Manual), LLC and VRM power delivery stability can account for a few degrees.

It would be interesting to know what workload differences you see between P95 Small FFT's versions in terms of power consumption in Watts on your 9700K. I'm presently alternating between 26.6 and 29.4 default Small FFT's on a test rig, as I've done countless times, and they're within +/-1°C at 155 to 156 Watts package power with all Cores displaying steady workloads on the graph, as is typical and expected.

Also worth noting is that AVX offsets on some motherboards may not function properly in Auto, or in any manual setting, which can be model specific or affected by BIOS versions.

Do you have any opinion or suggestion with regards to what and how I QC new builds and/or any suggestions to optimize the process?

I'd like to reiterate that as per Intel’s Datasheets, TDP and Thermal Specifications are validated “without AVX”.

See page 87, Section 5.1.1 Thermal Considerations, first paragraph, second sentence - 8th and 9th Generation Intel® Core™ Processor Families Datasheet, Volume 1

Accordingly, I never use 29.4 for determining thermal performance; it's a tool for making power, thermal and leakage current comparisons such as between lidded versus delidded, or for determining AVX offset stability. Even if users run AVX apps such as for rendering or transcoding, those real-world workloads seldom approach that of pre-AVX Prime95 v26.6 Small FFT's. Nevertheless, even if users don't run AVX apps, it's recommended that an AVX offset of at least 2 or perhaps 3 be configured to assure stability, in the event that AVX workloads may be used at some point.

CT 😎
 
  • Like
Reactions: cholzman

Karadjgne

Titan
Ambassador
Memory controller. It's inside the cpu. Small fft doesn't use much ram, so the mc doesn't contribute much to cpu temp. Blend and large fft use considerable if not all available ram, so the mc does contribute to cpu temps. But thats also based on the assumption that each version is running the exact same test so blend in 26.6 can be different to blend in 29.4 and show different results. With 26.6 built for ddr3 and 29.4 built for ddr4, I can't see it using the exact same test process even if the procedure is the same. Or there would be no need for newer versions if it was one size fits all programming.
 
  • Like
Reactions: cholzman

cholzman

Prominent
Aug 25, 2018
8
2
515
I want to thank everyone for your replies. I will admit some of this is going over my head and I feel the extremist testing is not something I condone (such as delidding, etc...). However, the information offered sets me on the path I was looking for so I can better understand a reasonable testing process for an average customer.

Consider this my virtual handshake and sincere appreciation for taking the time to provide me with this technical information!

Carey
 

CompuTronix

Intel Master
Moderator
Prime95's purpose is to find prime numbers .. by any means possible. Versions later than 26.6 added support for AVX, then AVX2 and FMA3 Instruction Sets followed by AVX512, each of which accelerate mathematical calculations and increase workload on the FPU, which in turn increases Core temperatures.

Coincidentally, the Prime95 pre-AVX version 26.6 Small FFT's is also ideally suited for the purpose of testing CPU thermal performance. It's better suited than any other test in any other non-proprietary utility, because it's a steady-state 100% TDP workload which yields steady Core temperatures.

Intel describes their test conditions in the datasheets, which includes a non-AVX steady 100% TDP workload to validate thermal specifications. The conditions further specify ambient temperature, all BIOS settings at Intel (not motherboard) default values, all DIMM slots populated, and motherboard on an open bench with the stock cooler.

Since TDP (Thermal Design Power) ratings are derived from the fastest hyperthreaded CPUs per generation and family (also called "flagship" processors), only those CPUs will reach TDP at stock settings. For example, although the 4 Core 4 Thread i5 7600K follows the same design and shares the same 91 Watt TDP rating as the 4 Core 8 Thread i7 7700K (flagship), the i5 won't reach TDP as it has a lower frequency and lacks hyperthreading. This is also true of the 95 Watt i5 8600K / i7 8700K as well as the 95 Watt i7 9700K / i9 9900K. So all hardware, software and environmental variables being equal, higher frequency and / or hyperthreading produce higher Core temperatures at the same TDP.

CT 😎
 
  • Like
Reactions: cholzman

Karadjgne

Titan
Ambassador
For me, games don't use AVX, or if they do, it's such small amounts that it's really negligible. They also don't use AVX2 or FMA3 and especially not AVX512. So running p95 26.6 is essentially simulating the same conditions and instructions that gaming would, to a 100% load. If the cooler can handle p95 26.6 at 100% and stay relatively close to 70°C or under, it'll handle any game the user can throw at it, whether it's gta:v on a 4thread cpu at 100% usage or the Sims at 45% usage.
 

cholzman

Prominent
Aug 25, 2018
8
2
515
I have no idea how my customers use their PCs. Some may game, some may do video editing, some may do nothing more than browse the web and email. Others may do all three.