Omega Thermocouple (in IHS) Vs. Onboard Sensors

Accidently posted this in wampeters...

Please Read the following.. You will now understand why I so emphatically state that the onboard sensors are totaly useless.



Test System
DFI nF4 Ulta-D w/ Vcore, Vdimm and Vtt mod
AMD A64 3000+ Winchester with T-type thermocouple mod to IHS (lapped IHS)
2x512mb OCZ PC4800 EL plat
ATI X800GTO
Omega HH506RA Multilogger Thermometer w/2 T-type thermocouples

Cooling System
Cooltechnica MP-05 SP Pro
Danger Den Maze4 GPU
Cooltechnica AquaExtreme 50Z 12VDC Pump
Danger Den Black Ice Extreme 2
2x Yate Loon 120mm fans @12v
Daul 5inch Bay Res
Clearflex 60 1/2" ID 3/4"OD tubing
Loop Order - Pump -> GPU -> CPU -> Radiator -> Reservoir


Test Method
Ambient temp taken with 36gague T-type thermocouple mounted in front of intake fan of radiator
Actual CPU temp taken with 36gauge T-type thermocouple embedded in CPU IHS with Omega bond OB-101 thermal epoxy
CPU Load created with S&M v1.7.6 beta with FPU test
Temperatures were allowed to settle for approximately 30 minutes before readings were taken.




Thermocouple

Ambient
---------------------
23.2 +-0.1C

Idle
---------------------
27.7C (with 22.7 ambient)

Load
---------------------
32.2C

dT (CPU Load to Ambient)
---------------------
9.0C



Onboard Sensor

Idle
---------------------
27-29C (ambient between 22.7-23.2)

Load
---------------------
40 +-1C (jumps around)

Ha, 8C diff.

yup.. just goes to show you how useless they are.

Unless, of course, the onboard sensor was actually inside the CPU core.

Thanks for the infor. MaxxxRacer.

So the actual Load temp's much lower than the mobo report.

indeed it is much lower..

From idle Ive noticed about a 4-5C jump in temps when loading the cpu.. this is independent of ambient temp.. Not the 12C jump that the onboard sensors show.

Overconfidence: the onboard amd sensor is IN the cpu... if you setup a proper calibration curve and use a Maxim thermistor sensor + SM bus reader, you can actually get accurate results from the onboard sensor... but its a pain in the arse and requires modding your mobo and cpu.

Quote:

---

Originally Posted by MaxxxRacer

Overconfidence: the onboard amd sensor is IN the cpu... if you setup a proper calibration curve and use a Maxim thermistor sensor + SM bus reader, you can actually get accurate results from the onboard sensor... but its a pain in the arse and requires modding your mobo and cpu.

---

But.. technically... the temperatures would be slightly higher on the inner sensor, wouldn't they? (Not the 8C that it's suggesting right now, but something?)

well I will admit that the IHS temp is a bit insulated from the core, but the IHS temp provides an average temp of the cpu... About 70% (guestimate here) of the cpus heat goes directly upwards, so where the probe is, most of the energy will be transfered, thus providing a good analysis of the heat being outputted..

anyway back to what you asked. the core temp will be a litttle higher (in some spots on the core) than the temp reported by the IHS.. but here is the thing.. the onboard sensor is placed on the edge of the core where it is the coolest... I even asked an AMD engineer about this and this was the info he provided.. he also poitned out that in this section of the CPU (on A64 anyway) the temp is a bit lower than the rest of the core...

So if the temp where the sensor is is lower than the rest of the core and the sensor is reporting 8C above what I am reading with the IHS temp... well u can see where this is going.

MaxxxRacer, is this appiled to IHS less cpu as well ?

i'm asking because of i'm running opty170, IHS off, my load temp, according to MBM5 is about 4-5 C higher than idle temp. (idle ~ 28-29C, load ~ 32-33C ).

the temp difference between load and idle? No idea what an IHS less cpu will do compared to an IHS capped cpu.. with the method of temperature measurement that I am using it is impossible to mount the sensor to an IHS' less cpu..

BUT your idle and load temps do look alot more realistic than most... Although i should point out that the 3.5-5C delta I demonstrated was with a 2.25ghz 1.573v winchester will not be the same as your opteron 170 which puts out alot more heat especailly if its overclocked.. (although the output could be less than i am assuming if your running like 1.2v or something rediculous like that.

I'd like to see this on a DFI Expert. My Smartguardian on my latest Expert which is an older revision reports temps below ambient during idle.

Nice job maxxx :fact:
Thanks for the info, I just stressed to view the temperature of my processor was fluctuating. :nono:

OMG ! Now I understand how much the onboard sensors suck.

Quote:

---

Originally Posted by Coroner Kyle

I'd like to see this on a DFI Expert. My Smartguardian on my latest Expert which is an older revision reports temps below ambient during idle.

---

so send me an expert ;)

While I understand that on-board temperature sensors (OBTS) suck, it yould really be nice if they scaled linearly. It appears as thought the OBTS is off by a percentage of the actual temp, making it difficult and almost useless to setup an offset which some temp monitoring programs allow you do.

You can't even compare your own before and after temps with any modicum of reliability..... what a bummer.

One thing I did notice though.. the idle temp of the cpu follows the actual temp relativley closely, even with changing ambients. my ambient gets to around 12C at night if i leave the window open.. when its cool in my room the cpu will idle in the low to mid 20's.. the onboard sensor reflects this as well, and my load temps are a bit lower as well. BUT with the lower the ambient temp, the worse the accuracy of the load temp.. for instance when i ran S&M when it was about 14-15C ambient, the differential between actual cpu temp and ITE was about 10.0C... earlier today when ambient was 22.4 the difference between actual and ITE was lower...

Hi Maxxx :) . Make such statement as "just goes to show you how useless they are." is absolutly unadapted. It isn't worth to make such tests if you start arguing with that kind of statement without knowning what you do, what you measure and what is the influence of whole parts below your thermocouple. You are testing with pre-fabricated ideas about sensors capabilities (like a lot of people) and this is not correct ! Yes, temperatures reported by sensors (and more exactly by the chip) can have an offset due to traces on PCB, noise, etc., but generally, evolution is fine (diode inside die is good and known with a certain tolerance, remember yourself the Phaestus calibration of 1700+, it was linear).

Thermocouple T° and die T° give 2 interesting informations. Obviously, the die is hotter than thermocouple because we need to take in account thermal resistances of the IHS (quasi negligible here), of TIM1 (~0.04-0.07°C/W) and of the silicon die (often based on 0.03-0.04°C/W). The sum of these thermal resistances isn't small at all, mainly on AMD with their poor TIM1. There's nothing strange to get an 8°C difference between thermocouple and cpu here, on the contrary, it's realistic and your die T° is quite good in regards of dissipated power. Please, stop thinking that a die don't warm, it's false (hot spots, etc.)... You have no idea of what a TIM (and others parts) influences thermal gradient. One proof (need more controlled data and studies cases but it's a beginning) is that at very low power at Idle, thermocouple T° and die T° are in the same range and close, because T° gradient occasionned by previous thermal resistances is small (low power->low gradient). The better will be at 0 power to see the real offset, but we need to use external circuitry to control internal diode, not easy. When power raises, this gradient raises too linerarly (pay attention to secondary losses) : the more power generated, the more gradient is and the more the difference between thermcouple and die will be. You need to control environnement around mobo, a big change in ambiant will change influence of secondary losses (draught?thermal paste settled? other parameter?), it's not necessary linear and it's probably not... Need a lot of testing to understand thermal mechanisms and influences here to define an acceptable way of testing.

Roscal

I agree with rosco here. The information provided thus far is fairly meaningless, sorry Maxxx. I know it is just a start (hopefully) but it is going to take lots and lots of data and tests and of critical importance is the control of ambient temps. What I will say, is that this is the sort of temp measurement I prefer, seeing the die temp and the IHS temp instead of just one or the other. Anyway, hopefully this is just a start and hey, you got Roscal to register :)

rosco.. interesting info.. and not to sound like an arse.. all of which i know (probably from reading your posts are procooling).


guys, maybe you didnt hear, but the ambient temps during the load testing didnt vary but by .1C. If you guys think it would be beneficial I could add another thermocouple infront of the radiator fans?

"The information provided thus far is fairly meaningless"

how is it meaningless, and how would you go about making it more meaningfull.

I suppose the thread should have been entitled "Omega thermocouple vs. Motherboard" because, as rosco pointed out, the on die sensor, when attached to a proper reading source, is surprisingly linear.

One other thing I forgot to mention in the test setup. I am using artic lumiere which, according to the arctic silver guys, settles in all of 15-30 minutes.

In any case, I appreciate the criticisms (despite however defensive I may get), and would greatly appreciate any ideas you guys might have on how to improve on the test methods.


On a side note: I was just thinking to myself about the calibration methods derek used on the 1700 and A64 chip... they were done with no power going through the chip... what happens when you put power through the chip? does the calibration still hold true, or is it effected by the various signals in te chip?

I ask this question not to attack Derek or his methods (infact I tried the same method and learn a good portion of what I know about testing directly from him), but rather as an honest question.

If the on die snesor is effected by the chip running or not running, that does indeed pose an interesting challenge. and on the same note.. Is the thermocouuple I have embeded in the IHS affected by the chip running too? and if so, how would we ever go about overcoming such obstacles?

Could you not cut a slot into the heatsink and mount a temp sensor in this ? That way you could run without the IHS, and if you mount the sensor correctly you will get the temp at the surface of the die.

Just my 2 cents.

Maxxx, I should have stated that the results thus far are kind of meaningless. I think it is common sense (?) that the die temp would be higher than the IHS temp. The data that will be interesting to me, is the IHS to die temp difference among different water blocks. I still think these 2 will not scale in a linear fashion. Bill keeps saying that die temps are meaningless and tcase is where it is at. How could having both sets of data hurt?

Maxx > Sorry, I read again the entire topic and I have to complete my answer above, because I forgot something and I was misleading myself (reading too fast probably). I thought you used water T° to get the deltaT, but you used only ambiant air sucked at radiator. This induces some variation generally, and mostly you have a GPU WB in the loop -> you mix all of them and you get an unbelievable disorder in your data at final...

What do you want to catch exactly? Whole system perf or only WB perf? Because it's not the same. I read others threads and I saw you want to get WB performances principally, it's OK, but do we need a GPU in the loop to do that? No (and if it's only a first try, no abusive conclusions should be taken as above). The loop should contain the minimum to avoid undesired variance (you don't control GPU dissipated power for example). Here, you are introducing some factors which are not necessary to establish the CPU WB behaviour (without forget flow for future data). You should take the water T° just at inlet of the WB, it's more precise and consistent (physicals parameters need to be as close as possible of the things you want to measure, always). T° in front of a radiator can vary a bit because of fan recirculation at edges and hub dead zone, so multiples thermocouples are generally recommended in each quadrant, but here we don't need them, water T° is far easier to get. You'll have a waterchiller I believe, but it's not really necessary too. It's a facility to get faster results because we don't have to wait a long equilibrium establishement with radiator. But if you decide to keep you water, let's say @30°C +/- 1°C (not important to get 30.00°C, you won't see any difference at 31°C in your C/W curves because water properties don't change radically within 1-2°C), you can play with fan regulation slowly to keep water in your range, it's not difficult when air T° is quasi constant because you can adjust precisely the fan before reaching equilibrium (a first try give some infos about final T°). I do that, just a habit.

About diode calibration, I think there's no interaction with others CPU parts because it's just a simple diode packed into the die to measure current response through it. The first idea in my mind is to verify it (or at least have an idea) with a massive underclock to get low dissipated power at Idle, if it's OK (no interactions with power going though the chip or mobo connexions), you'll find again a water T° really close of your die T° (low power > low gradient > low T° difference). Good luck for your tests.

PS: Is your first sentence ("[..]like an arse[..]") a criticism ? I don't understand well the meaning...

XS forum is ultraaaaaa slow at this time, isn't it?

rosco, this was merely a brief snipet to show the oddities of the onboard sensor..

for any testing that will be for reviews i will follow a much more error hardned method of testing... Chiller instead of radiator, waterblock inlet and outlet temps (water temps recorded with T type thermocouples), with the inlet temp held constant with the haake chiller.. ambient temp also recorded and the cpu will be on a dedciated loop powered by an iwaki WMD20RLT.

niksub1: it is common sense that the die temp will be higher than the IHS temp, but the amount is a bit much i think, and the variation between die and ihs temp with varying ambient temps.. for any testing, the sensor in the IHS is fine, even if it may not be exactly with the core temp, becuase it will be a comparison between different waterblocks with the same testing methods... (I hope that made sense). But just for you I will include the core temps for all data points, however useles I think it is. lol.

Tcase IIRC is the ihs temp where i am measuring (been a few weeks since i looked at an intel tech doc with the names of the measurements points). If so, then I would have to disgree with bill.. the actual core temp is what we REALLy want to know... while Tcase is very useful and perfectly fine for the kind of testing i will be doing, if you want the absolute best conception of how exactly something is cooling the cpu, the core temp is where its at. but with that said, the on die sensor is poorly placed for such operations..


the like an arse comment was not a criticism aimed at you. I was merely pointing out that what i was about to say may come off as rude, or snide, and it was not my intention.

while I dont have a calibrated way of measuring the temperature of the die (sold my maxim kit to fund my current goings on), i will undervolt and under:banana::banana::banana::banana: the cpu to its lowest levels tonight and re-run the tests.. If i have time i will also remove the gpu from the loop.

Quote:

---

Originally Posted by MaxxxRacer

niksub1: it is common sense that the die temp will be higher than the IHS temp, but the amount is a bit much i think, and the variation between die and ihs temp with varying ambient temps.. for any testing, the sensor in the IHS is fine, even if it may not be exactly with the core temp, becuase it will be a comparison between different waterblocks with the same testing methods... (I hope that made sense). But just for you I will include the core temps for all data points, however useles I think it is. lol.

---

Oh but it is not useless. This was my whole argument against TTV testing... all that is shown and known is Tcase. I am interested in how the 2 scale together (if at all) with respect to different blocks. Will there ALWAYS be an 8C difference between core and IHS? Will that gap widen or shorten depending on the block? All very interesting and not useless at all.