Intel TAT / CoreTemp / IDCC all different temperatures....

[unclewebb]

I originally posted this at the [H]ard forum but I think it's relevant and this procedure can be used to test how accurately programs that people take for granted actually track the DTS.

I use CrystalCPUID to read the on chip model specific register ( MSR ) that contains the raw DTS information. The 7 bits of temperature data are located in bits [22..16] of MSR 0x19C in all Core based processors.

Here's my E6400 ( 3600 MHz 1.504 volts ) at idle in a 22C room with the Tuniq Tower fan at its lowest speed of 1150 rpm.



When you enter 0x19C in the MSR Number box and click on RDMSR ( Read MSR ) you will get the value of the DTS. In this example bits [22..16] corresponds to 36 hexadecimal which is ( 3 X 16 + 6 ) = 54 in decimal. This means at idle I am 54C degrees away from TjMax which is the Intel documented highest safe operating temperature for a C2D cpu.

If RMClock is using the DTS and calculates the absolute core temperature to be 33C then it must be assuming that my TjMax = 33C + 54C = 87C. That's certainly possible but now let's check it again at full Orthos load.



At 3600 MHz and 1.480 volts it is happily running Orthos within 18C of TjMax. RMClock reports a core temperature of 63C so now it must be assuming that TjMax = 63C + 18C = 81C.

The maximum junction temperature of your processor ( TjMax ) is a fixed value, set at the factory and doesn't change yet RMClock is using two different values depending on whether the processor is hot or cold. That's not right. I also performed this test with TAT and the calculated TjMax could change from day to day which is also impossible.

Now you can run CoreTemp 0.95 or SpeedFan 4.32 and you will see that both of these programs track the on chip digital thermal sensors exactly. CoreTemp may be guessing wrong at TjMax but it can be used to report the DTS directly. When it is set like that, it reports the temperature data in processor register 0x19C exactly.

This is the only Intel documented core temperature information for a Core 2 Duo desktop processor.

Edit: On the main page of CrystalCPUID you can set it to read any of your cores. TAT is using different TjMax values at idle, at full load and even CPU 0 and CPU 1 are not consistent.

[joebuffalo]

This is the only Intel documented core temperature information for a Core 2 Duo desktop processor.

...just though it deserved to be repeated...

[SLi_dog]

heh heh heh, apologies for digging out this old thread but it's just an invaluable resource for Core temperature/DTS monitoring info


Before I start I should make it clear I understand that all Core2 desktop CPUs have no Tjunction value and as a result they should probably all be configured to read the DTS values however there is already a massive amount of info out there in regards to the Conroe Core 1066Mhz FSB Core2Duo and Kentsfield Core 1066Mhz FSB Core2Quads that have a "safe" CoreTemp load threshold of 65C and 75C respectively. Changing that perception, imho, would take more work than any benefit it may achieve.




As I'm sure you're all aware, the new 1333FSB processors are another CPU that seems to be having incorrect readout issues with the normal configuration of CoreTemp. I also have a sneaking suspicion that this applies to the E2140/E2160 CPUs as well.

This is all well and good as all who have followed this thread know you can just configure version CoreTemp v0.95 to report the DTS readings.

The problem comes with new users and what appears to be a lack of an easy to find thread which details the processors that are affected by this Tjunction issue and also a detailed method on how to configure CoreTemp to show the DTS.


I admit my understanding of actual CPUs affected by this is limited compared to many who post here, but I am assuming that:

E4300/E4400/E4500
E2140/E2160
E6x50 (1333FSB based CPUs)

Should all have their core temperature monitored by reading the DTS.


Another general assumption I'm working with is that the Core2 CPUs should be run no hotter than within 20C of the DTS 0C mark.
- This assumption is based on the first Conroe CPUs where initial testing seemed to show that staying under 65C (20C lower than the 85C nominated Tjuction) was the best threshold for a "safe" overclock.


I have also put together this image on how to configure CoreTemp v0.95 to show the DTS readouts.





This kind of information would make a handy sticky in the "Intel Info And Support" section for easy reference to new users of Core2Duo/Core2Quad.


I can start a thread if people here think it's worthwhile but I'd prefer to get some feedback on whether the CPUs listed above are the main ones affected by this issue.

There's also need to be a Mod to sticky it otherwise it's pointless.

[bing]

Doing it the correct and Intel's way DOES REALLY matter !

Ever wonder how are those Intel CPU engineers measure Intel's own CPU in their uber leet lab ?

Warning, cpu pron pics inside, not advisable for minor !

Intel® Core™2 Extreme Quad-Core Processor and Intel® Core™2 Quad Processor Thermal and Mechanical Design Guidelines Download the PDF file.

For those who likes quick shortcut, just go to page 81 and beyond !

After you've read that and if you believe it, next time you will start chuckling & gigling back there when you hear these kinda comments :

"TAT cpu temp reading rocks ! I ignore others"
"Coretemp is the best and I believe it !"
"Temp diode reading is the most accurate one"
"The Tjunction should be xxx, not yyy .... "
..bla..bla..

For those non believers even after you've read it, c'mon, if this Intel's document can not be trusted, then what else ?

Or that program that shows the temp that you "want" to believe ?

Conclusion ? Software based for absolute CPU temp reading is NEVER ACCURATE, period !

Edit:Repost.

[unclewebb]

SLi_dog: Another general assumption I'm working with is that the Core2 CPUs should be run no hotter than within 20C of the DTS 0C mark.

The general assumption that DTS=20 or 65C is a good number to shoot for when overclocking is too vague. 65C is an undocumented absolute temperature and people's definition of overclocking varies. My E6400 runs fine at 80C (DTS=5) when overclocked to 3200 MHz but 60C (DTS=25) is a good upper limit when I'm pushing it to 3600 MHz and a little beyond, God willing.

Software based for absolute CPU temp reading is NEVER ACCURATE, period !

That's my conclusion and it applies to all Core processors. There's a reason why Intel engineers outline such an elaborate process to come up with an accurate Tcase temperature.

The new revision G0 processors, especially the dual cores, use significantly less watts at idle than the previous B2 processors ( 8W vs 22W ) when C1E is enabled. That's even less than the 12 watts that the L2 processors ( E4x00 ) are rated at. This is resulting in some impossible looking idle temps that are well below ambient when air cooled. For me, it's just more proof at how inaccurate the DTS is at reporting low idle temps which it was not designed for. I believe the DTS may start losing accuracy below 50C so even load temps for well cooled G0 dual core processors at default MHz and voltage may not be accurate.

Using the DTS to report absolute core temperatures the way SpeedFan and CoreTemp do is not documented by Intel and is just plain wrong. No matter what TjMax is chosen, I don't believe you can get 100% accurate core temperatures this way across the full range of temperatures that a Core processor can operate at.

The DTS was designed and calibrated to signal and control the thermal throttling and shut down points of an Intel Core processor and for that purpose, it works excellent. Using CoreTemp 0.95 as shown above to keep track of temperatures relative to the throttling point is the only thing users should be keeping track of and using the DTS for.

[SLi_dog]

The general assumption that DTS=20 or 65C is a good number to shoot for when overclocking is too vague. 65C is an undocumented absolute temperature and people's definition of overclocking varies. My E6400 runs fine at 80C (DTS=5) when overclocked to 3200 MHz but 60C (DTS=25) is a good upper limit when I'm pushing it to 3600 MHz and a little beyond, God willing.

It may appear to be more vague that it really is. By definition, running your CPU no closer than 20C from the DTS 0C point means exactly that. You'd have to factor in maximum CPU load and maximum ambient operating temperatures into that threshold. eg, don't go setting the 20C DTS rule as an idle temp at 6am in the middle of winter

I kinda threw it out there as a rough value for beginner users but it is well and truely open for discussion and I actually wouldn't apply it myself in the case of a Core2Quad. I guess the question then is, if not 20C, then what? Maybe 25C as you use?

Also, I personally wouldn't run a CPU closer than 20C to the 0C DTS point at 24/7 settings regardless of whether it was at 2.4Ghz or 3.4Ghz so I may have misunderstood what you're saying there.

Conclusion ? Software based for absolute CPU temp reading is NEVER ACCURATE, period !

That's my conclusion and it applies to all Core processors. There's a reason why Intel engineers outline such an elaborate process to come up with an accurate Tcase temperature.

Absolutely and that's fine for you, Bing, myself and everyone else who has followed this thread but I guess you're missing the point of my post . My point was that this information needs to be easily accessible to the average overclocker and it doesn't appear to be.

The new revision G0 processors, especially the dual cores, use significantly less watts at idle than the previous B2 processors ( 8W vs 22W ) when C1E is enabled. That's even less than the 12 watts that the L2 processors ( E4x00 ) are rated at. This is resulting in some impossible looking idle temps that are well below ambient when air cooled. For me, it's just more proof at how inaccurate the DTS is at reporting low idle temps which it was not designed for. I believe the DTS may start losing accuracy below 50C so even load temps for well cooled G0 dual core processors at default MHz and voltage may not be accurate.

This seems to contradict your previous statement agreeing that software based absolute temperature are inaccurate. If they're inaccurate for absolute temperatures, I can't understand how they can be used to indicate that the CPU temps are lower than ambient?
Admittedly my understanding of thermodynamics is reasonably limited but I find it hard to believe that any amount of current could exist at a lower temperature than the air that's being used to cool it (Note: I understand about surface area and heat transfer)


At the risk of repeating myself, I guess you're missing the point of my post. My point was that this information should be easily accessible to the average overclocker. From the large amount of threads on this subject here and at other online OC forums, it currently doesn't seem to be.
From following this thread I assumed that there was enough know-how and experience here to put something together that could be of great use to the wider OC community.

[bing]

Absolutely and that's fine for you, Bing, myself and everyone else who has followed this thread but I guess you're missing the point of my post .

My point was that this information needs to be easily accessible to the average overclocker and it doesn't appear to be.

Agree, if its accurate, it will be every OC-er dream isn't it ?

But unfortunately the answer is "NO", you won't and can't get that info since Intel doesn't have it for them self !

1st thing pop out from the mind after reading my above statement is "Huh ? Why & How the heck I know that ?"

Ok, here are my explanations, its just basic electronic that every EE learned on their 1st year and "common sense".

1st, you have to understand that popular "tdiode" in that silicon die is basically a plain jane diode, and it is the very basic structure of silicon electronic.

A diode consist of "PN junction" (try wiki yourself, since it will be too long here) that has an interesting physical characteristic, it has a predictable and very linear behaviour response on it's temperature coefficient which is approx. 1.22 mV/deg F or 2.2 mV/deg C.

So whenever there is a changes in the temperature, we could track the changes (delta) temp quite accurate, hence it is used intensively to control the cooling parameter in Intel CPU mechanism, for example adjusting fan speed accodingly, or throttling up/down the PWM controller and etc. The word PECI come in mind !

Now you might ask how about the absolute temp reading ?

This is the weakest point, cause no diode in this world have a same temperature offset even they were produced in the same batch. Unless you could make that diode by assembling atom by atom at the precise location for all diodes.

So to make an accurate absolute reading, we must calibrate each diode in a controlled temperature room at high & low temp to get the offset temp for "EACH" diode.

If Intel want to give us an accurate temp reading on "EVERY TDIODE" in the CPU they sell, they must calibrate "EVERY TDIODES" that is spreaded in that CPU at silicon die level. Fyi, each CPU has more than one Tdiodes.

Layman term, they have to toast every cpu die and measure every tdiode in it at the temperature controlled room at cold and hot in order to get the offset of each tdiode and write/burn down that offset value into the CPU die.

Sounds complicated isn' it ? Ok, even it is not complicated, you must agree that this "extra" process needs an extra production cost right ?

Now, imagine you were Intel management, that always struggling to optimize your production cost at each CPU. Do you want to do that ?

Remember, Intel "officially" never endorse or encourage overclocking on their CPU.


Hope this helps.