e8400 Core Temp: Data Analysis Confirms it is Largely Meaningless

[ixtapalapaquetl]

It should be stated from the outset that my chip seems to run very hot, even relative to other e8400's, and my conclusions may not generalize. In my quest to bring temps down I have removed the IHS (no help) and installed fairly high end water cooling (some relief). However, you will likely agree the temperatures that follow are still quite high. I could not rationalize how my system could possibly find stability given these temps - reason told me that I should crash as Core Temp exceeded 95C.

Earlier today I was messing around in Everest and found that one can set it to log readings in Preferences and store this info in a .csv file. Messing around, I ran Prime off and on a couple times so I would have idle and load data. This generated something like 1281 data points. I opened the file in Excel and started playing.



It is interesting to note in looking at the second load phase that you can almost see the transitions from 1024k to 8k to 10k etc. These CPU temps are the ones I have been relying on given the numerous complaints about Core Temp values and my own experience. This got me thinking... I have tons of data I could use to explore this anomaly.



The first thing I noticed was the marked difference in the (ahem) differences of the two measures at idle versus under load. Clearly there is a more pronounced difference between the idle temp readings. This did not strike me as something that makes sense - if both CPU and Core Temp purport to measure the same value (ie heat emanating from the CPU), I would expect the two to be aligned across the entire range of values.

I started manipulating columns and came up with the results above. The difference between CPU Temp and Core Temp at Idle is over twice that under load. Note the standard deviations - the idle values are much more consistent than the load values. So we see two things going on: at idle, Core Temp is much greater than CPU Temp, but consistently so; and at load, Core Temp is only marginally greater than CPU Temp, but has a much greater relative fluctuation.

I began to strongly suspect that Core Temp is not related to CPU Temp in any way nearly as meaningful as we would naively believe it to be. This led me to consider the correlation between the two sets of values.




I found this information to be fairly damning. The correlation coefficients tell the tale - r^2 values of .4744 and .657 respectively indicate only a very weak correlation between CPU Temp and Core Temp. If these were both measures of heat emanating from the CPU, I would strongly expect the correlations to be well above .9. The fact that they are not suggests to me that one of these two measures do not mean what we think they do. I am inclined to believe that CPU Temp is our good old fashioned, well, CPU Temp, whereas Core Temp is a measure of something only peripherally related to CPU temperatures.

So that's that. I attached the raw data in case you'd like to confirm any of my calculations, and I encourage you to use Everest to generate your own. I would love to know if other e8400 users find similar discrepencies, and would also be very interested in seeing if 65nm chips behave in the same manner.

[Note: I had to change the extension to .txt to upload it; rename it .csv or .xls to use it in a spreadsheet. Also, the first graph should read 5 second intervals.]

[Shampoo]

Very interesting read.
I remember back in the days of overclocking Athlon XP processors (Barton core) that Core temperatures weren't available because there was no thermal diode in the core and everyone overclocked based on CPU temps.

Never really had a problem there did we?

Then came the core temperature readings of more recent cpus and everyone started taking the cpu temp lightly, compared to the core temp measurements

Intel also states that you should take the core temps with a grain of salt right?

Lots of effort put into this analysis of cpu temperatures.

Very much appreciated.

[TheKarmakazi]

This is very nicely presented and quite informative.

I'm going to start using that feature of Everest, since I feel that its the most accurate temp readings from the programs I have used.

Shampoo: I remember hearing that the core temps may not be 100% accurate also

[tomb18]

CPU Temp is the temperature measured by a motherboard sensor located near the CPU. It is calibrated to give you a value similar to the Tcase, or the case temperature.
The Core temperature is about 15C higher which is pretty well what you are seeing above.
If your CPU is rated at a Tcase of 65C, the core temperature will read 80C which is fine.

[Leeghoofd]

I think there are many borked 45nm CPU's out there no matter what software proggy you rurn the readouts are really a disaster... I have one qx9650 that idles at 45°C at stock speeds ( those are CPU temps yes not coretemps ) and goes up to 59°C when being used... no matter what software proggy I use to have a readout... If I push the CPU to 4Ghz readouts stay the same lol...this is on the P5K DLx and Maximus ( both watercooled with different setups...) I really think Intel has some work here as something is seriously out of bounce... asI cannot believe those values I read out are correct, way off there....

[RPGWiZaRD]

Why do you even concider thinking the software temps are correct for these 45nm chips? I've barely looked at my temperatures, for me it's enough if it's OCCT stable (if temps were really high it would certainly crash at 4GHz 1.31 ~ 1.33v) and putting my hand at the exhaust Tuniq Tower 120 the air is barely any warm (well it's more like mild at best), A LOT colder than the air that was exhausted from E6750 3.75GHz @ 1.52v which core temp showed around ~60C full load temp during winter if that's somewhat accurate this E8400 should be more like 50C lol. I'm tired of all this whining about high temperatures on E8400 across the Internet, they aren't they're cool as ice, it's just software sensors are misleading.

[Leeghoofd]

My point exactly PRGWIZARD, all the theories in the world might be good for that CPU , another user might not get good values... something is pretty wrong here... it's just to bad for the user that we don't have much reference to work on regarding CPU temps anymore...

[ixtapalapaquetl]

CPU Temp is the temperature measured by a motherboard sensor located near the CPU. It is calibrated to give you a value similar to the Tcase, or the case temperature.
The Core temperature is about 15C higher which is pretty well what you are seeing above.
If your CPU is rated at a Tcase of 65C, the core temperature will read 80C which is fine.

The data shows that this is indeed not the case. The relation between the two is clearly not linear (ie add 15C).

Why do you even concider thinking the software temps are correct for these 45nm chips? I've barely looked at my temperatures, for me it's enough if it's OCCT stable (if temps were really high it would certainly crash at 4GHz 1.31 ~ 1.33v) and putting my hand at the exhaust Tuniq Tower 120 the air is barely any warm (well it's more like mild), A LOT colder than the air that was exhausted from E6750 3.75GHz @ 1.52v which core temp showed around ~60C full load temp during winter. I'm tired of all this whining about high temperatures on E8400 across the Internet, they aren't they're cool as ice, it's just software sensors a misleading.

I agree with your sentiment about testing stability and not "whining" about temps; I have no choice but to focus on stability. However, excessive temperatures is not the point of this thread. Instead, it is merely to point out that Core Temp does not seem to refer to the same information as CPU Temp. This information is new to me, and hence the thread.

...it's just to bad for the user that we don't have much reference to work on regarding CPU temps anymore...

Maybe they don't matter anymore? If you are stable, you are stable. Perhaps you could follow RPGW's lead and ignore them. I know I do not flinch in the least at readings in the 90's with this particular CPU.

[cadaveca]

Yes, Core temp and cpu temp are not the same...they never have been. While your results are interesting, I find it hard to take for full value; the thread title lead me to beleive you had REAL data, not just data from one cpu. one cpu does not create fact, except for that particular cpu.

E8400 and temp reporting accuracy varies from cpu to cpu. Why, I don't know, and like you, I really could care less about temps.


But I will tell you this, temps may not affect cpu stability so much, at higher levels, but THEY WILL MAKE THE CPU'S LIFE MUCH SHORTER. "They" being high temps, of course.


So, get about 100 or more cpu's run the same tests, correlate, then your thread title would be accurate.

Maybe look into WHERE on the cpu the CORETEMP sensors are...you may just find that this is the HOTTEST part of the cpu, while cpu temps are related to overall temp of the silicon die/dies.

remember, that cpu temp is a single value for quads...but there are two wholey seperate pieces of silicon. So, for quads, cpu temp reports what? Again, overall temp, not the hottest part of each die.

[jonny_ftm]

I second cadaveca too.
Your results are not that abnormal. CPU temperature is anlog reading somewhere on the die, between the cores. Removing your IHS also suerely makes your CPU readings no more reproductible with other CPUs.

It is just normal, that the more you increase the load/heat on cores, th emore they get hot while CPU temperature will increase in a slower manner. This is just normal physics, as the source of the heat is the cores where DTS sensors are, while CPU temperature is far away from the cores.

Personally, I'm not shocked by your results. Even without removing the IHS, the linear relation between coretemp (deep temp) and IHS temp (superficial temp) is lost on high OC.

But, other reports on DTS sensors being freezed is more than abnormal.

What you could do is use a temp sensor and put it on one core, and another on the CPU die. You'll see that they confirm the non linear relation

[ixtapalapaquetl]

@cadaveca, johnny_ftm

First and last sentences from first post:

It should be stated from the outset that my chip seems to run very hot, even relative to other e8400's, and my conclusions may not generalize.

I would love to know if other e8400 users find similar discrepencies, and would also be very interested in seeing if 65nm chips behave in the same manner.

I do not disagree with either of you. But if the above statements were unclear, allow me to try once again: I am simply requesting that others who are interested in the odd behavior in 8400 temperatures perform a similar analysis so that we might see if this phenomenon extends beyond one CPU (as I suspect it might).

Procedure:

1) In Everest, File->Preferences->Logging
2) Check "Log sensor readings to CSV log file:" and choose destination directory.
3) Check Time, CPU utilization, CPU Temp, and Core #1/Core #2 Temp.
4) Run at idle and under load long enough to generate significant data. It may be easiest to Prime overnight (at high overclock) - set it and forget it.

5) Open .csv log with Excel.
6) Sort by CPU utilization column.
7) Create a new column: Core Temp entry - CPU Temp entry.
8) Using this new column, find averages, standard deviations of 100% usage and non-100% usage for CPU Temp and Core Temp. Post results.

9) Create a scatter plot of CPU Temp vs. Core Temp for 100% usage. Include trendlines, correlation. Repeat for non-100% usage. Post results.

As you pointed out, this might very well be an isolated case. And it might not. I strongly suspect that we will not see the same behavior in 65nm CPU's as we do from these 8400's. Unfortunately I have but one CPU, and thus cannot investigate further without assistance. If my suspicion bears true, we will possess solid information to begin theorizing on causes. If not, we are merely left in exactly the same position we are now - namely, speculating without data.

[jonny_ftm]

I could try it, I'll see...

But, I still can't understand your questionning. It is well established, since the Core2 begining (quiet 2 years now) that there's no good correlation between CPU temp and Core temp. So, it is normal you won't find any correlation in your analysis. The reasons are many, like I explained above (sensor technology, position, Internal Heatsink...)

Also, all trials to understand how those coretemp sensors work ended to a fail. In fact, Intel never published the full spec datasheets on those DTS sensors. All programs run on a guess and try based on many CPU samples and comparisions between ambiant and coretemp on idle, low volt and under clocked CPUs.
The thermal throttle point was prooved to be more correlated to core temp than CPU temp.

So, yes, don't expect to find a correlation between those values, that's really well known. That's all I'm trying to explain.

Now, the 45nm DTS sensors being different than on 65nm, is something I never read about. It is too early to state it as most 45nm CPUs, especially quadcores aren't out yet.

I will see if I can do the stats you ask for, as, in any case, I don't expect to see any correlation