How to make sure your CoreTemp / TAT readings are right

[bing]

Here, hope it helps :





Also go HERE to have better "overall" picture in order to help you to understand few facts such as :

1. There is NO reference what so ever referring to what you guys called max. Tjunction absolute value.
2. Why the delta temperature is accurate.
3. What the reason behind the scene that Intel don't/can't provide an absolute cpu temp feature.

Ok, have to admit it that for points 2 & 3, you have to understand a bit of what is called basic PN silicon junction temperature cooeficient.

OT, watch for SpeedFan, Alfredo (the author) will release the latest version that has two options in reading the Intel cpu temp (currently avail in beta) which are :

1. The right & correct method using is relative temp to max. tjunction as above.
2. The absolute cpu temp with has "adjustable" max. tjunction, so it will be suitable for people who want their cpu temp looks bragging cool since it can be adjusted until it reads what ever the owner want to see. LOL

 

[Notlag]

Here, hope it helps :

...

Also go HERE to have better "overall" picture in order to help you to understand few facts such as :

1. There is NO reference what so ever referring to what you guys called max. Tjunction absolute value.
2. Why the delta temperature is accurate.
3. What the reason behind the scene that Intel don't/can't provide an absolute cpu temp feature.

Ok, have to admit it that for points 2 & 3, you have to understand a bit of what is called basic PN silicon junction temperature cooeficient.

OT, watch for SpeedFan, Alfredo (the author) will release the latest version that has two options in reading the Intel cpu temp (currently avail in beta) which are :

1. The right & correct method using is relative temp to max. tjunction as above.
2. The absolute cpu temp with has "adjustable" max. tjunction, so it will be suitable for people who want their cpu temp looks bragging cool since it can be adjusted until it reads what ever the owner want to see. LOL

Thanks for the excellent post, Bing, and the attachments (which will hopefully clear some things up for folks). It is possible to get a pretty decent idea of Tjmax (or how much to offset CoreTemp readings) via the method in the OP (the two listed as being more precise - not the loose estimate) because you know the actual core temps will be reasonably close to ambient (or generally at least within 2C or so) at idle w/ the processor clocked as low as it will go and the voltage as low as it will go. Then you can compare those temps w/ ambient.

But.... That's assuming you have a decent way of measuring ambient temperature (people would be surprised how many thermostats and thermometers have questionable accuracy) or one that you consider "good enough". It's also assuming that the temps read from the cores are at least linear and accurate, which may be questionable (I don't recall the exact reference, but I seem to remember some docs which suggested that the accuracy is better at the extremes, but not so much in the middle - don't quote me on that though).

Still, adjusting for it by estimating Tjmax is a far far cry better than taking the absolute temps given by TAT or CoreTemp as truth. The more important thing is making sure you're happy w/ your delta value If it says you're 20C, 30C, 40C, etc... away from Tjmax, what are you content with and what do you consider safe? It's going to vary from person to person, but generally people do have some absolute temps in mind before hand that they'd like to see.

 

[bing]

Why there is no accurate absolute temp reading from CPU ?

It's also assuming that the temps read from the cores are at least linear and accurate, which may be questionable

Yes, it is quite accurate for the delta and imo "very linear".

Basically the silicon thermometer is made from a single PN junction aka diode aka the most basic/fundamental semiconductor structure that builds up a CPU.

Layman, it is like a piece of brick with a bloop of cement to your whole house/building.

Its called forward voltage temperature coefficient at constant current, which is very predictable and almost linear against the temperature, but the problem is the offset which needs a calibration at "every" PN junction or "every" thermometer.

This will not be a problem "IF" Intel capable of laying out the PN junction "atom by atom" at the precise same location and it's crystal structure for all diode/PN junction/thermometer at every CPU they made.

For example try HERE and start reading at the "Calibrating Your Diodes" section, you will get the idea how easy to have a really good & accurate temperature delta, but its sucks and difficult to get the absolute temperature, unless you toast the thermometer at two known temperature points and draw the slope.

You can find tons of similiar pages like I pointed above from Google, or even at our own OC front page HERE.

If you read the white paper at my previous post, you will understand why it is almost impossible to have a single temperature reading that can represent "properly" for overall cpu temperature. Fyi, Intel embedded multiple thermometers throughout the silicon die.

Intel or could be AMD as well just simply run every single CPU up to certain level during the binning process, and when certain weak spot at the silicon die started to error at certain temperature from a particular temp sensor that near that weak spot, then they just marked that particular temperature point (aka the max. tjunction) or sort of burn that marking into that particular cpu, hence every CPU might have different max. tjunction point.

This is the major reason and also it is pointless & too costly for CPU maker to make an accurate absolute reading from CPU since it is not needed.

Again, the word "accurate temp" here means accurate at the delta, NOT absolute temperature. People have been misused this word "accurate" when it comes to Intel cpu temp discussion cause lack of understanding.

Now, once you understand this whole picture, imo Intel doesn't care heck at what absolute temp that point is. Don't you agree ?

If it says you're 20C, 30C, 40C, etc... away from Tjmax, what are you content with and what do you consider safe? It's going to vary from person to person, but generally people do have some absolute temps in mind before hand that they'd like to see.

Agree, its up to personal preference.

IMO, for an OC-er, I believe 20C or even 10C to Tjmax is quite safe even for 24/7 if it is not the wall for the OC, and that is because for the headroom say like if your A/C is broken or summer hot temp.

Why ?

Try hang out at computer forums "around the world" and watch tons of threads that sounds something like this ->

"Wow ! Just realized all this time almost for "years", my gf's rig was throttling like hell, no wonder it was sooo sloowwww. That is because of that damn noob at that our local PC store was not putting the cpu HSF properly. Once I fixed it, it works like champion."

Got my point ? Those CPU can take quite some beating.

OT, but still related, my other tip -> Finding CPU Throttling Temp sometimes will boost your OC morale !, especially for those who loves & believes absolute temp reading from software.

 

[Notlag]

It's not a question of whether or not the absolute temps will be an accurate measure of the overall temperature of the chip. They're not. They never will be - there are going to be areas of the chip which are hotter than others based on which portions are active at any given time.

The question is whether or not it is possible to convert a temp given by TAT or CoreTemp, which is actually relative to thermal management kicking in, and convert that to an absolute temp. You've shown the linearity of thermal diodes, so provided one can determine Tjunction, then one can convert the temperature from a relative temperature to an absolute temperature. The only thing in question is the value of Tjunction, and that can be estimated.

Consider the following reasonable approximation:

Absolute Temp = Power Dissipated * Thermal Resistance of Cooling Solution + Tjunction + Temperature Delta

or...

Tjunction = Absolute Temp - Temperature Delta - Power Dissipated * Thermal Resistance of Cooling Solution

The heat sink's thermal resistance, as well as the thermal resistance of the die / IHS interface, the thermal resistance of the TIM, etc... are fairly hard to accurately estimate, but if you reduce your processor's voltage and speed you can decrease the amount of power dissipated to the point where the product of it and the thermal resistance is small enough that the error in the estimate doesn't contribute much to the overall equation.

At that point, assuming the temperature delta reading from the CPU is accurate relative to Tjunction, then you can estimate Tjunction as accurately as you can measure the temperature of the air going through the heat sink fan (give or take any remaining error in your estimation of the power dissipation and thermal resistance).

If I run my E4300 at 600 MHz and 0.875v, then at idle it dissipates approximately 2W. At worst, even with a terrible heatsink, a CPU w/ a bad die / IHS interface, mediocre mounting, and crummy TIM, the sum of thermal resistances might be 0.6 C/W. And it is much more likely to be around 0.3C/W to 0.4C/W or less total with realistic parts. That means the cores at that speed should be around 0.6C over the air going through the heatsink fan. Even if I'm terrible at estimating such things, I should still be able to safely estimate Tjunction within about 1C of my ability to measure the temperature of the air passing through the heatsink fan. That said, most thermometers are reasonably repeatable, but often not very accurate, so measurements of the ambient air temperature may be a couple degrees off. In all that is about +/-3C potential error.

That much error may sound like a lot to many people at first glance, but it's MUCH more realistic than the temps people take at face value today every time they look at Core Temp or TAT. How many people have you seen at this forum claiming their U120X is cooling their cores to at or below 20C at idle even when overclocked? Their house would have to be quite brisk.

Yes, when it comes to determining how far you can push your processor, the delta is much more useful, but in the end people still want absolute temps for talking with or comparing to others. Are the adjusted absolute temps really more meaningful? No. But they are much more meaningful than the absolute temps given by TAT or CoreTemp w/o adjustment.

The other good thing about this method is that they can do it safely. Determining the throttling temperature by actually causing throttling is riskier than determining throttling by reducing temperatures to ambient (admittadly, MANY CPU's have ran throttled for extended periods w/ no meaningful damage or decrease in useful life span, and given the way technology becomes obsolete, a CPU's life span isn't really that important).

EDIT: Btw, very interesting reading in the links you sent Bing.