You are not logged in.

#1 2009-06-26 09:17:55

new2arch
Member
Registered: 2008-02-25
Posts: 235

lm_sensors / coretemp accuracy

Hi guys,

I'm wondering how accurate the core temperature readings are with lm_sensors?

Yesterday, with ambient room temperatures hitting ~28-29 C, idling core temps rose to 45 & 48 degrees C after couple of hours of inactivity.
That can't be right!
My system consists of following items:

Case: Antec P182,
Mobo:Asus P5N-E SLI,
Processor: Intel C2D 2.0GHZ @1 Volt (undervolted from stock 1.32) w/speedstep,
RAM: 2GB Corsair DDR2 RAM,
GPU: Asus Geforce 7300GS,
System Fan: Scythe S-Flex 120mm @ ~600rpm.
CPU-fan Nexus 92mm @ 850 rpm.
CPU-heatsink: Ninja Mini
Arch - kernel 2.6.30-5
lm_sensors -  3.1.1-1

The Ninja Mini's heatpipes close to its base are not warm to touch at all - Perhaps normal body temperature was more like it.
Wouldn't I have noticed if the idle core temps were hitting close to 50 C?
BIOS report CPU temperature (T-case?) being 38-39 C.
Exhaust temp (near system fan) is 36 C
Running Orthos for 30 minutes under WINE gives me 56 & 57 C on both cores. And, again, the CPU has been undervolted!
The Ninja Mini heatsink has been properly installed with a tiny blob of Arctic Cooling MX 2.

Temperature inside the case (measured with a analog thermometer lying on the bottom of the case) was 30-31 degrees C.

I'm beginning to believe the thermal diodes aren't either perfectly calibrated or the analog -> digital conversion is giving off digits.
And /or the lm_sensors in Linux aren't giving the correct values. Or a mix of all three?

Quote:

"(C) Tcase and Tjunction are both acquired from Thermal Diodes. Tcase and Tjunction analog to digital (A to D) conversions are executed by separate devices in different locations. BIOS Calibrations from motherboard manufacturers, Factory Calibrations from Intel, and popular temperature utilities are frequently inaccurate."

"(E) Intel provides only partial documentation for Tjunction Max on desktop processors. For Throttling and thermal Shutdown protection, Intel uses the Digital Thermal Sensors (DTS) to monitor Delta to Tjunction Max, which is a relative value that varies from Core to Core, and is not an absolute temperature."

and...


"(F) Tjunction Max must be known to calculate absolute Core temperature, which is Tjunction. Popular temperature monitoring utilities may incorrectly estimate undisclosed Tjunction Max values, which results in excessive Core temperatures and inconsistent Tcase to Tjunction Gradients among Ci7 and C2 variants."

Unquote

http://www.tomshardware.co.uk/forum/pag … _11_0.html

According to many reports, it seems getting accurate core temps is not very easy.

With lm_sensors in Linux, I noticed the critical limit for cpu temperatures are way over the E4400's max (T-junction?) temps of 61 degrees C or thereabouts stated by Intel.  I don't recall those maximum numbers with lm_sensors right now but with lm_sensors they're closer to 90 C or even higher.

Perhaps this has something to do with the core temperatures being inaccurately reported?
I've re-seated the CPU heatsink several times, hoping that any inconsistency of the heatsink's base's flatness and/or the Intel heatspreader's dito would be remedied but I see only very small changes in temps.

Any ideas?

EDIT:

It seems the tj_max has to be known or else the core temp readings will be inaccurate. Perhaps I'm misunderstanding this discussion at Kerneltrap:

http://kerneltrap.org/mailarchive/linux … 804/thread

Last edited by new2arch (2009-06-26 09:34:33)

Offline

#2 2009-06-26 10:12:55

panosk
Member
From: Athens, Greece
Registered: 2008-10-29
Posts: 241

Re: lm_sensors / coretemp accuracy

new2arch wrote:

Hi guys,

I'm wondering how accurate the core temperature readings are with lm_sensors?

Yesterday, with ambient room temperatures hitting ~28-29 C, idling core temps rose to 45 & 48 degrees C after couple of hours of inactivity.
That can't be right!

Why not? These values seem very reasonable to me.

BIOS report CPU temperature (T-case?) being 38-39 C.
Exhaust temp (near system fan) is 36 C
Running Orthos for 30 minutes under WINE gives me 56 & 57 C on both cores. And, again, the CPU has been undervolted!

I think these values justify my opinion. Since you have 38-39 in BIOS, where the system has almost zero stretch, +5-8 degrees with a fully loaded desktop is not much.
Also, I think 56-57 C while gaming (I suppose Orthos is a game, right? big_smile) is a perfect temp. Just for reference, my Dell XPS 1330 with KDE idles at 40-42, and reaches 60-70 under heavy stretch (gaming with World of Padman or compiling), which I consider very reasonable.
In any case, I am not saying that lm_sensors or any other temperature measuring utility is absolutely accurate, but they definitely serve their purpose.

Offline

#3 2009-06-26 10:26:59

new2arch
Member
Registered: 2008-02-25
Posts: 235

Re: lm_sensors / coretemp accuracy

panosk wrote:

Why not? These values seem very reasonable to me.

Hello,
Lot of folks with similar 65 nm dual 2 cores are idling at 30-35 degrees C or even much less, just couple of degrees higher than ambient.
I don't think near 50 C is a normal /good idling temperature.


panosk wrote:

I think these values justify my opinion. Since you have 38-39 in BIOS, where the system has almost zero stretch, +5-8 degrees with a fully loaded desktop is not much.
Also, I think 56-57 C while gaming (I suppose Orthos is a game, right? big_smile) is a perfect temp. Just for reference, my Dell XPS 1330 with KDE idles at 40-42, and reaches 60-70 under heavy stretch (gaming with World of Padman or compiling), which I consider very reasonable.
In any case, I am not saying that lm_sensors or any other temperature measuring utility is absolutely accurate, but they definitely serve their purpose.

In BIOS, processor is 100% burdened. Temperature reported by BIOS is more likely T-case temp which is totally different from T-junction...and T-case is always couple of degrees lower than T-junction.
Orthos is a cpu burning test that will produce a controlled, "synthetic" worse case scenario where all cores (and memory if you wish) are put at full work for as long as you wish.

You idle 40-42 with lm_sensors? I don't know the TDP specs for your cpu, or how well the speedstep works or how good the airflow is etc. For notebooks, I think the core temps are expected to be higher than desktop pc's due to various factors involved, mainly because the former usually/always come with a lower performance heatsink, tiny fans, and bad airflow compared to the latter. 
If your cpu were producing similar levels of heat as mine, you'd expect that my core temps would be much lower than yours, since I have a larger and "better" case, better cpu-cooler, bigger fans etc etc.

Last edited by new2arch (2009-06-26 10:28:26)

Offline

#4 2009-06-26 10:50:49

panosk
Member
From: Athens, Greece
Registered: 2008-10-29
Posts: 241

Re: lm_sensors / coretemp accuracy

Hello,

Well, you have definitely searched on the subject more than me, so I can't argue :-). The truth is I don't bother with temps, unless I see sth very weird. I have a nice plasmoid that shows various hardware info. Few weeks ago, core temps skyrocketed to nearly 80 C while the lap was idle. However, the cpu was running at 100%, so the problem was apparently a process. Indeed, a plasmoid went crazy, so after uninstalling it everything was normal again. Have you checked the cpu load and the processes in case sth works a little harder than usual? I suggest this because I still think the temps you get from the stretching utility are normal (for a desktop machine).

Cheers

Offline

#5 2009-06-26 11:02:10

new2arch
Member
Registered: 2008-02-25
Posts: 235

Re: lm_sensors / coretemp accuracy

panosk wrote:

Hello,

Well, you have definitely searched on the subject more than me, so I can't argue :-). The truth is I don't bother with temps, unless I see sth very weird. I have a nice plasmoid that shows various hardware info. Few weeks ago, core temps skyrocketed to nearly 80 C while the lap was idle. However, the cpu was running at 100%, so the problem was apparently a process. Indeed, a plasmoid went crazy, so after uninstalling it everything was normal again. Have you checked the cpu load and the processes in case sth works a little harder than usual? I suggest this because I still think the temps you get from the stretching utility are normal (for a desktop machine).

Cheers

I should probably take your advise and just ignore the reported temps...
That's the easy "solution" but I'm not 100% happy with it.
If someone could confirm that lm_sensors often gives off wrong reports, I'd be happier. :-)
My cpu load at idle are ~3-7%, both cores.

80 C is certainly way too high and if not throttled it will shorten the chip's life span.

Edit -
As a side note, my southbridge is ~50 degrees C and it's pretty hot to touch. No way the Ninja's heatpipes could feel ambient or body temperature cool, if the processor was giving off nearly as much heat as the southbridge. And what's the idle watt consumption for dual 2 core 4400? Less than 10 W.
The Ninja's fins are cool to touch as well.

Last edited by new2arch (2009-06-26 11:09:01)

Offline

#6 2009-06-26 11:12:23

Crows
Member
From: Wales
Registered: 2008-09-05
Posts: 92

Re: lm_sensors / coretemp accuracy

I couldn't tell you if lm_sensors' figures are reliable for your model but 80 is much higher than your CPU should ever get. You seem to know your way around the inside of a case - anything like dust or bad seating that could cause heat like that should be pretty obvious. If you can't see any of that stuff, it's a safe bet that the sensor is wrong.

PS. Dell XPS M1330 here too. Thermal spec for the T8100 is over a 100 degrees. Should probably stop getting nervous when it hits 60.

Offline

#7 2009-06-26 11:16:09

new2arch
Member
Registered: 2008-02-25
Posts: 235

Re: lm_sensors / coretemp accuracy

Crows wrote:

I couldn't tell you if lm_sensors' figures are reliable for your model but 80 is much higher than your CPU should ever get. You seem to know your way around the inside of a case - anything like dust or bad seating that could cause heat like that should be pretty obvious. If you can't see any of that stuff, it's a safe bet that the sensor is wrong.

PS. Dell XPS M1330 here too. Thermal spec for the T8100 is over a 100 degrees. Should probably stop getting nervous when it hits 60.

Not direclty related to your post, but I've found some Arch user's reporting the inaccuracy of lm_sensors:

http://bbs.archlinux.org/viewtopic.php?id=49442

Pretty old post though, and I don't know if the new kernel + lm_sensors would affect the print outs.
The recent kernel + lm_sensors didn't affect my core temps.


Edit:

Here's the t-junction max according to lm_sensors:

sensors (idle temps, ~20 minutes of inactivity) :

CPU Temperature:    +35.0°C  (high = +90.0°C, crit = +125.0°C) 
MB Temperature:     +41.0°C  (high = +70.0°C, crit = +125.0°C)

Core 0:      +43.0°C  (high = +74.0°C, crit = +100.0°C)
Core 1:      +45.0°C  (high = +74.0°C, crit = +100.0°C)

I also tried the Intel stock heatsink (note: it belongs to E4700 and it's thinner than the stock HS for E4400) to see what it'd do to the temps -
The idle core temps went up to 50/51 within 1-2 minutes and stayed there.
Fan was controlled by Asus Q-fan, it didn't ramp up until I ran Orthos. Within 1 minute of Orthos, core temps hit 60 C.
At that point, I quit running the software.

I doubt the stock HS (okay, it belongs to the E4700, which I assume is slightly cooler than mine given its thinner appearance compared to the stock HS for E4400) wouldn't manage to keep my E4400 @ 1 volt within safe limits. It doesn't make any sense that the cores would reach maximum temps within 1-2 minutes even with the stock HS - normal computer usage with stock HS should allow both cores to stress for longer periods of time than that. I'm inclined to believe the lm_sensors is giving the wrong core temps.

Last edited by new2arch (2009-06-26 16:54:17)

Offline

#8 2009-06-30 16:03:50

new2arch
Member
Registered: 2008-02-25
Posts: 235

Re: lm_sensors / coretemp accuracy

*Major update*

So I took a spare HD and installed W2K on it.
Downloaded Orthos (newest version), Real temp (version 2.5 and 3.00), HWmonitor (newest version) and Core Temp.

Idling these are the temps I'm getting (ambient room temp: 28 degrees C):

HWmonitor (no indication of Tj-max, but I believe it's 100 C. *T-junction max is set as 100C. This has been verifired by increasing the T-junction max in RealTemp 3.00, and then I get same numbers as HWmonitor):
Core 0 - 50 degrees C
Core 1 - 52 "

Real Temp 2.5 (t-junction max setting is 85 C):
Core 0 - 30 C
Core 1 - 32 C
Distance to TJ-max: 55 C, 53 C.

Core Temp: (TJ-max: 95 C)
Core 0 - 42
Core 1 - 44


Stress test:
A session of Orthos 15 minutes, Large FFT etc:

HWmonitor:
Core 0 - 57 C
Core 1 -58 C

Real Temp 3.00 (T-junction max setting is 90 C in this version):
Core 0 - 47
Core 1 - 48

Real Temp 2.5 (TJ-max - 85C)
Core 0 - 42
Core 1 -43
Distance to Tj-max: 44, 43

Core Temp (Tj-max 95 C)
Core 1 - 52
Core 2 - 53


These numbers are weird. Don't know which one to trust. They are all guessing the TJ-max values thus spitting out different results... the TJ-max value varies from 85-95-100 degrees C!
I also tried running orthos with Vcore Auto [~1.32 V] set in BIOS, 5 minutes = ~ 10 degrees C over previous load temps @ 1 volt!
This demonstrates how valuable undervolting really is.

Lm_sensors in Linux most likely has around 100 degrees C set as TJ-max (in spite of it saying 74 degrees C "high" and "critical" 100 C) *this has also been verified by increasing TJ-max in RealTemp 3.00 to 100 C and I get the same output as Lm_Sensors), so it's unlikely that my temperature readings are accurate...I've read reports concerning the E4400 chip's TJ-max being less than 90 C. So, either Real Temp 2.5 or 3.00 is giving the correct numbers AFIWB (As Far as I Want to Believe). I could probably slice off at least 10 degrees or even 15 degrees from the output I'm seeing in lm_sensors.

So...I think my idle temps are 35 & 37 degrees C if TJ-max is assumed being 90 degrees C for my chip. Moreover, it doesn't make any sense that the cores are idling at 2-4 degrees over ambient temps (Real Temp 2.5) with aircooling.


Any ideas?

Edit 1: minor misspellings
Edit 2: added my hasted conclusions
Edit 3: did a new test to fill in some blanks

Last edited by new2arch (2009-06-30 19:33:08)

Offline

#9 2009-07-29 13:26:02

vit
Member
Registered: 2008-12-28
Posts: 79

Re: lm_sensors / coretemp accuracy

I think you should contact with lm_sensors developers and show them the situation.

Offline

Board footer

Powered by FluxBB