ED NOTE: This is a reprint from Mike’s site HERE.
My main goal is to provide extra knowledge into Socket A temperature measurement problems. As of today, numerous “big” and reputable websites, such as Anandtech and Tom’s Hardware, have published very poorly done heatsink “comparisons”. In this article are reasons why these big websites should not have published their results, and why they should have tried to find a better solution to the temperature measurement problem.
My personal opinion is that big websites have a responsibility to their viewers to provide accurate results to their tests. Using the excuse of “the test was repeated twice” does not suffice for socket a temps, because the incorrect/compressed results are indeed repeatable.
As a side note, I have attempted to contact most websites that have done incorrect reviews. For the most part, the response is a cold-shoulder response. On other occasions, they have replied with the “I did the tests twice, so they have to be correct” response. To me, I do not believe either answer is appropriate. To me personally, their unwillingness to change their reviews is one sign that these websites truly do not care about the end-user.
Personally, I do not have the time or the resources to create my own solution. One site that I do think is really striving to find “truth” is 2CoolTek. I feel that his “newest” test method of using an insulated 72W peltier / heatsink testing setup will probably be the best thing to come as of yet as far as testing/reviewing (and far better than the non-relative inaccurate socket A mobo tests that exist on the web).
Acknowledgements: Nevin House, John Carcich (both heavily responsible for my limited knowledge of thermodynamics), Joe Citarella of Overclockers.com for posting numerous new tests into the accuracy and relativity of Socket A temps, Stephen Hoar of Burning-Issues and the countless others who I have tried to learn from.
Problem I: Measuring from a Secondary Heat Pathway
Every different motherboard has slightly different thermistor placement location. For example, the ABIT KT7 touches the backside (underside of the chip core edge), whereas the AMD760 based Motherboard that Toms Hardware uses is surrounded by resistors (which give off heat) and does not touch the backside of the CPU. So the underlying problem with comparing temps between these two mobo platforms is in the fact that the two motherboards do not measure temperature at the same location.
The further problem with backside temp measurement is that the measurement of CPU temperature is based on a secondary heat pathway. Secondary Heat Pathways are usually non-isolated temperature measurements, at locations where the temperature is significantly different than that at the primary heat pathway (CPU Core to Heatsink). Looking at an this Overclockers.com article, where they use a "isolated" thermocouple placed directly behind the CPU core surrounded by Closed Cell Foam in an effort to isolate the thermocouple, one can see a higher-temp reading than the KT7 thermistor (which contacts backside CPU core) by roughly 25%. So this secondary heat pathway from which the KT7 measures exhibits at least a 25% loss compared to the primary heat pathway, and in most instances likely higher due to most KT7s not isolating the installed thermistor.
Update : AMD states that thermal resistance between Ccore to Cback (core to back of core) is 0.5. So on a KT7, where the drop from directly behind the core to core edge (where the mobo measures temp) is around 25%, a KT7 registers less than half of the CORE temp change. And then there are the ORB-style heatsinks, whose airflow interacts with socket-thermistor temps, further distorting the temp readings (all the orbs (C-ORB, Super-Orb, Arcticooler) benefit in this manner when tested on a Socket A platform. On these heatsinks, the temperature measured is even further less of a percentage of core temp change. Since there are instances of 6X compression, there likely is a range of compression for socket-thermistors from 2x compression all the way to 6x compression.
Problem II: Thermistor is not isolated from air contact, preventing a any accurate guess work from the known resistance from Ccore to Cback
Having examined the secondary pathway temp loss problem (not in detail,though), backside thermistors show only a fraction (less than half, probably something more like 35-30% on a KT7) of what the actual CPU core temp change is, which in turn leads to both Inaccurate Results and Inaccurate, Non-Relative
Comparisons. Assuming that the KT7 is one of the more accurate mobos as far as temps go, others would appear to be more inaccurate, especially the ones that do not touch the CPU backside.
On top of this, the thermistor in almost none of these motherboards are isolated from air contact. Of the motherboards with thermistors that contact CPU backside, only a portion of the thermistor actually touches the CPU, with the rest contacting (and further disturbing the temp reading) Socket A. And in most instances, air is free to move in and out of the backside socket.
Problem III: Compression of Temp results due to measuring a secondary heat pathway that is neither isolated from outside influences "air" or with a known resistance from measuring spot to CPU core.
One sign that temps are compressed with Socket A mobos is the small difference in temperatures from Idle to Full Load, in most cases. For example, on the KT7, on most occasions the temperature change from idle to full load is only 5C. Comparing this idle drop versus a P3 internal diode reading (while itself inaccurate, it does not compress readings and does show full temperature changes) in the following graph:
Ambient Temperature was 22C. The heatsink used for both tests was the Taisol CEK733092 (~.45 C/W). Assuming accuracy for comparison purposes on a P3, the CPU wattage at idle would be roughly 15 W. Assuming accuracy for the Socket A setup, one would get at idle 50 watts.
Not to mention, that in order for the T-Bird to have a peak core temp of 42C, it would require an 13C ambient case temp.
Now, with the next graph, theoretical versus shown T-Bird temps will be compared. Since AMD chips do not offer the same HLT optimizations, I am guessing that a 1 GHz T-Bird will idle at 45W, with a peak of 60W.
A theoretical difference of 10C based on C/W calculations (with C = Temp of CPU – Ambient Case Temp) is read, by thermistor, as a 5C temp, which is a compression of temp readings as compared to CPU core temps.
The Taisol CEK733092 is measured at roughly .48 C/W. The ORB series of heatsinks is roughly .52 C/W (as determined by Burning-Issues). Based on Alpha specifications, the PAL6035 is roughly .38 C/W, depending on the fan used.
As can be determined by the thermistor readings on the Anandtech "comparison", the Alpha PAL6035 has a "read" temp of 45C, while the C-ORB has a measured temp of 47C. IF these readings were relative, then the theoretical numbers would measure out to be exactly the same temperature difference as in the review.
Taking the .37 C/W for the Alpha (Alpha rates their heatsinks with the 21 cfm Sanyo-Denki Fan), and .52 C/W for the C-ORB as measured by Burning-Issues, the theoretical temps for a CPU that creates about 55 watts of waste heat (as determined with the program
Radiate) would result in an actual core temperature difference is about 10C, not the 2C measured difference. Here is just one example of why the comparisons using Socket A platforms are not "relative" and do not give a good indication of which heatsink is "really" better. Case temp was revealed as 29C in this Anandtech BBS Thread. In addition, even between the two PAL6035s (one with a YS Tech fan, the other with a Sanyo Denki Fan), the 1C measured temp difference is more like a 3C difference.
One important item to note is that the Theoretical Temperature would be absolute best case scenario. Any number of factors would alter this theoretical number upward.
Now looking at the Toms Hardware Review, You can see a "measured" temp difference between pal6035 versus Taisol CEK733092 as 1C.
You can see the same general problems as the Anandtech review, and some new ones as well. The Taisol versus Alpha measured difference was 1C, but theoretically, that difference is more like 10C. The Alpha versus C-ORB was measured as 3C, but theoretically is roughly 8C. The new element is the unpredictability factor of the socket-thermistor temp measurements. For one heatsink, it may read 5C too low, for another 10C too low – all on the same mobo. Yet one more reason not to use Socket A motherboards as comparison platforms.
Unfortunately, the Anandtech review was "challenged" in their BBS. For various reason’s on the BBS thread, the review was not amended in any way to show that the comparison was highly unpredictable and inaccurate.
So you get concerns about both accuracy of these numbers, but the readily viewable "compression" of temp readings (mainly due to the fact that thermistor measures a secondary heat pathway that is a fraction of actual CPU core temp change) and an unpredictability factored into this "compression". In addition, due to compression and unpredictability, the thermistor read results are not Relative. Repeatable, definitely. Are they relative to actual core temp changes? Definitely not. These are the overriding factors that SHOULD keep websites of any size from posing reviews/comparisons with Socket A heatsinks. It is imperative to those larger websites, like Anandtech and Toms Hardware, who have far-reaching audiences, from posting comparisons that are both inaccurate and have results that are not even relative to CPU core temp changes.
Problem IV: Comparisons should not be done between separate end-users
I believe, that with the inaccuracies of socket-thermistor temp measurements, that they should solely be used as an "approximation" for any single user. Comparing temps between different systems and case setups is hardly feasible, as temps compared on a single motherboard are already impossible to do (with socket-thermistors). One thing you can do is to go to my solutions page and approximate your own CPU temp, and try to figure out how "inaccurate" the thermistor is for your particular temperature/voltage/overclock (it changes for each overclock, due to compression of temp changes).
Problem V: High Rate of Instances where the temp measured is highly inaccurate
There are instances of 10C or even greater accuracy errors. For instance, in the Anandtech "Comparison", there is a 12C accuracy error on the C-ORB, for an error of over 23%. On the UL BIOS as well, so even though the newer, compensated BIOS does a good job approximating temp, they are still approximating and are subject to potential huge reading problems. Then, in this recent Anandtech BBS thread, there is a user that is reading a 30C CPU temp, with a 23C system temp on a KT7. This is probably an error from actual core temp of almost 17C, for an error of ~43%.
Results like these are not abnormal for socket-thermistor setups. Unfortunately, they can be extremely common. Sometimes the temp "error" is 5-6C, other times its 10C, and sometimes 15C. The problem is that there is no way to predict how "off" the thermistor reading will be on a wide-case basis. It has been done on an individual basis.
Problem VI: P3 Socket-Thermistor Temps versus P3 Internal Diode Temps
In this section, the general theme is to further compare socket-thermistor readings versus trusted (sometimes inaccurate) but RELATIVELY accurate (as far as measuring temp differences) means. The first set of tests is done with a slotket (with a hole in the socket, through which a thermistor was placed directly behind the CPU core) on a BX6 R2 (with modifications to correct diode readings).
One problem is that the percentage of CPU core temp change DECREASES as CPU core temp increases. This situation can only be worse on a socket A CPU, as they run much warmer than P3s. Again, a compression of the Socket Thermistor Temps versus Internal Diode Temps is evidenced. Going between 995 to 566, a diode temp changes roughly 11C, whereas the Thermistor only registers a 6C change. Likewise, the 16C change with PCTC is registered as a 9C thermistor change. Also, you can tell the tendency of the socket-thermistor to under-read temps (to report temps lower than actual core temps).
Keep in mind that these temps are done with a P3, which operates at much lower wattage than Durons/T-birds. To figure out wattage, use Radiate.
Conclusion (of a definite work in progress):
Websites have a responsibility to the reader to provide both accurate, and in the case of cooling, relative results in their reviews/comparisons. With AMD gaining steam very quickly in the Enthusiast computer user market, it is also exposing a lot of new users to overclocking. For these new users, reviews like those at HardOCP, Anandtech, and Toms Hardware serve as confusion. If Socket A temp measurements were inaccurate, but accurately portrayed CPU core temperature differences, this webpage would not exist. The problem is the thermistor solution isn’t relative: You can see a 10C Core Temp Difference Shown as a 1C difference in the Toms Hardware review, and a 10C difference shown as a 2C difference in the Anandtech Review.
It is not my place to force upon these websites to at least amend their reviews to simply state that the measurements are not really relative, nor accurate. But I do feel that a webmaster/owner has responsibility to spread correct knowledge, not to spread disinformation and sow confusion.
VII. Temperature Compression further explained.
This section will further explore the temperature compression problem in more detail.
The first bit of information comes from AMD. They state that Ccore (Core temp) to Cback (CPU backside) resistance is .5. This is directly behind the core, where mobos are not measuring. So even if we assume that the mobo manufacturer is using a table to "approximate" temps, they are going by a table that varies based on thermistor placement and airflow interference, that is highly variable in nature.
Then we look at the KT7 mobo in particular for further examples. Looking at where the thermistor is on the KT7, it is behind the CPU core edge (backside). Looking at Overclockers.com tests comparing an insulated thermocouple versus the existing thermistor (keep in mind that this is isolated, helping to minimize airflow interference that is present in most situations), the temperature drop is around 25-30% from directly behind the core.
So we get the KT7 thermistor reading, at best 33-37.5% of core temp change, ignoring airflow disruption. Since we can’t ignore airflow disruption, and that certain heatsinks disrupt airflow more than others, you get the variability in thermistor readings that prevents any table from actually providing accurate data.
Since looking at some online reviews (Anandtech one in particular) where you get a temperature change of 2C read, versus 10C theoretical (see Problem III), you get a compression factor of 5x. Which means, at best, that in this particular reading, the thermistor was only registering 20% of CPU core temp change. So I would personally say in the ideal situation, with an insulated thermistor reading, a single user could at best hope for a reading that is representing 50% of core temp change and hope that the mobo manufacturer uses a .5/1 table for displaying temps.
At worst, you’re looking at certain heatsinks (Arcticoolers, Super-ORBs, Chrome ORBs and the way they have air coming off all sides of the heatsink seem to be the worst at this) which exhibit even greater socket-thermistor interference than do normal heatsinks.
So, this is yet another instance of why there is temperature compression and unpredictability, and inaccuracy issues.
NOTE: How To Calculate Your Own Theoretical Core Temp
You can calculate the “theoretical” temps for yourself. I have included several heatsink’s ratings below.
Calculate CPU Watts with Radiate.
All cases assume a perfect CPU to Heatsink connection, one that requires no grease (this will serve as the “minimum” temp over ambient you can achieve with the stock heatsink).
Use this formula:
CPU Watts * Heatsink Rating C/W = Degrees C over Ambient that your CPU core should be running.
For example, a Duron running at 1000 MHz and 1.85 volts radiates 50.3 watts. With an Alpha PAL6035 and an ambient temp of 25 C, the CPU Core should run at 50.3 watts * 0.38 C/W = 19.1 C over ambient, or 44.1 C.
.48 C/W (PCTC removed, heatsink lapped)
.45 C/W (PCTC removed, heatsink lapped)
.38 C/W (PCTC removed, heatsink lapped)
.52C/W (PCTC removed, heatsink lapped)
|Thermaltake Super Orb||
.38 C/W (with Sanyo-Denki Fan, improves with higher CFM Fan)
.35 C/W (with Delta 38CFM Fan)