[lm-sensors] coretemp.ko -- Sandy Bridge: temperature readings "biased" based on power consumption?

[lm-sensors] coretemp.ko -- Sandy Bridge: temperature readings "biased" based on power consumption?

[Frantisek Rysanek]

Dear gentlemen,

I've just tried looking at the CPU temperature via coretemp.ko on 
some SandyBridge CPU's. I have a desktop-grade Core I3 at 3.3 GHz, 
and a mobile Core i5 at 2.5 GHz. When I flip the CPU from "pretty 
much idle" to "each core running a cpuburn instance", the temperature 
reported by coretemp *jumps* up in a second. On the desktop CPU, the 
jump is about 10*C. On the mobile CPU, the jump is about 25-30 *C !
It jumps up in a second after I launch the software load - and then 
it proceeds to gradually inch up, as the massive heatsink starts to 
warm up a bit (perhaps another degree C in twenty seconds to a 
minute).

Should I look for some thermal mischief in my system?
Or, do the SandyBridge CPU's indeed "precompensate" the temperature 
reading based on instantaneous power consumption?
I.e., is the "digital thermal sensor" more like a "fan control hint 
with a strong feed-forward component", rather than a half-decent 
thermometer?

I've noticed already on some recent 45nm Core2 CPU's (on a gigabyte 
motherboard) that the fan control feedback loop responds magically 
swiftly to me launching some software load.
I know that modern SuperIO chips (containing PWM fan control logic) 
can take the CPU temp value straight from the CPU (perhaps mediated 
by the south bridge), using a digital link called PECI. I also know a 
bit about the autonomous fan control algorithms implemented in chips 
by ITE and Winbond(Nuvoton) - I've played with these before.
If indeed the CPU's DTHERM sensor would "bump up the reading" in 
response to power consumption, that would explain the swift 
modulation of fan speed based on instantaneous CPU load...

Any comments are welcome :-)
And, thanks for the providing the blessed coretemp.ko in the Linux 
kernel.

Frank Rysanek


_______________________________________________
lm-sensors mailing list
lm-sensors@lm-sensors.org
http://lists.lm-sensors.org/mailman/listinfo/lm-sensors

Re: [lm-sensors] coretemp.ko -- Sandy Bridge: temperature readings "biased" based on power consumpti

[Guenter Roeck]

Hi Frantisek,

On Wed, Oct 31, 2012 at 12:01:08AM +0100, Frantisek Rysanek wrote:
> Dear gentlemen,
> 
> I've just tried looking at the CPU temperature via coretemp.ko on 
> some SandyBridge CPU's. I have a desktop-grade Core I3 at 3.3 GHz, 
> and a mobile Core i5 at 2.5 GHz. When I flip the CPU from "pretty 
> much idle" to "each core running a cpuburn instance", the temperature 
> reported by coretemp *jumps* up in a second. On the desktop CPU, the 
> jump is about 10*C. On the mobile CPU, the jump is about 25-30 *C !
> It jumps up in a second after I launch the software load - and then 
> it proceeds to gradually inch up, as the massive heatsink starts to 
> warm up a bit (perhaps another degree C in twenty seconds to a 
> minute).
> 
I think that is pretty normal. Keep in mind this is the temperature within
the chip core, not at the external package.

I have systems with i7-2600 and i7-3770K. Both run at 30C or less when idle.
The core temperature jumps by at least 10C as soon as I add load. If the
load is at 100% for a minute or so on all cores, the temperature increases
to around 80C, and the CPU fan speed increases to keep it there.
Nothing to be concerned about unless your fan control is not working.

> Should I look for some thermal mischief in my system?
> Or, do the SandyBridge CPU's indeed "precompensate" the temperature 
> reading based on instantaneous power consumption?

Not really. It is just that with higher power consumption, the CPU temperature
can jump significantly within a few seconds. After all, the power translates 
into produced heat, which has to go somewhere. Since the sensors are directly
at the core, you see the impact immediately.

I think that may be a bit different to older CPUs, which probably tend to run a
bit hotter even when idle, resulting in less temperature increase under load.
Or, in other words, Intel probably did a good job conserving power when it
is not needed with recent CPUs.

> I.e., is the "digital thermal sensor" more like a "fan control hint 
> with a strong feed-forward component", rather than a half-decent 
> thermometer?
> 
> I've noticed already on some recent 45nm Core2 CPU's (on a gigabyte 
> motherboard) that the fan control feedback loop responds magically 
> swiftly to me launching some software load.
> I know that modern SuperIO chips (containing PWM fan control logic) 
> can take the CPU temp value straight from the CPU (perhaps mediated 
> by the south bridge), using a digital link called PECI. I also know a 
> bit about the autonomous fan control algorithms implemented in chips 
> by ITE and Winbond(Nuvoton) - I've played with these before.
> If indeed the CPU's DTHERM sensor would "bump up the reading" in 
> response to power consumption, that would explain the swift 
> modulation of fan speed based on instantaneous CPU load...
> 
power -> heat -> increased temperature ... no need to be too fancy.

Hope this helps,

Guenter

_______________________________________________
lm-sensors mailing list
lm-sensors@lm-sensors.org
http://lists.lm-sensors.org/mailman/listinfo/lm-sensors

Re: [lm-sensors] coretemp.ko -- Sandy Bridge: temperature readings "biased" based on power consumpti

[Phil Pokorny]

[-- Attachment #1.1: Type: text/plain, Size: 1261 bytes --]

On Tue, Oct 30, 2012 at 5:32 PM, Guenter Roeck <linux@roeck-us.net> wrote:

> Hi Frantisek,
>
> On Wed, Oct 31, 2012 at 12:01:08AM +0100, Frantisek Rysanek wrote:
> > Dear gentlemen,
> >
> > I've just tried looking at the CPU temperature via coretemp.ko on
> > some SandyBridge CPU's. I have a desktop-grade Core I3 at 3.3 GHz,
> > and a mobile Core i5 at 2.5 GHz. When I flip the CPU from "pretty
> > much idle" to "each core running a cpuburn instance", the temperature
> > reported by coretemp *jumps* up in a second. On the desktop CPU, the
> > jump is about 10*C. On the mobile CPU, the jump is about 25-30 *C !
> > It jumps up in a second after I launch the software load - and then
> > it proceeds to gradually inch up, as the massive heatsink starts to
> > warm up a bit (perhaps another degree C in twenty seconds to a
> > minute).
> >
> I think that is pretty normal. Keep in mind this is the temperature within
> the chip core, not at the external package.
>

And there is internal thermal paste inside the CPU package between the CPU
die itself and the integrated heat spreader.  That thermal material also
explains why the silicon die temperature can jump so quickly.  It has a
higher deg C/W than the IHS and the heatsink on top of that.

Phil P.

[-- Attachment #1.2: Type: text/html, Size: 1664 bytes --]

[-- Attachment #2: Type: text/plain, Size: 153 bytes --]

_______________________________________________
lm-sensors mailing list
lm-sensors@lm-sensors.org
http://lists.lm-sensors.org/mailman/listinfo/lm-sensors

Re: [lm-sensors] coretemp.ko -- Sandy Bridge: temperature readings "biased" based on power consumpti

[rayer]

Hi,
yes Guenter is right. Intel placed the DTS sensor into hottest place 
inside CPU core(s).
This is needed for safe thermal shutdown if any problem. Due to thermal 
impedance
between silicon, IHS and heatsink it happen that heat cannot flow so 
fast through silicon
to heatsink (thermal conductivity is less than metal) and fast 
tepmperature rise will occur.
It's normal. On my C2D system with E8500 it has idle temp about 40C and 
when I run
CPU burn it will jump about 10C in few seconds then slower the rise up 
to 60-65C.
CPU fan doesn't react so fast on each load peak, it spin up slowly if 
longer load remains.
So I don't see any problem at all.
It may be more visible on latest IvyBridge CPU with poor chip to IHS paste.
It doesnt affect notebooks because they don't use IHS (heatsink directly 
on silicon die).
On desktop CPU it was proven that removing the IHS may take temp some tens C
down under load, esp. when overclocked.

Martin

Guenter Roeck wrote:
> Hi Frantisek,
>
> On Wed, Oct 31, 2012 at 12:01:08AM +0100, Frantisek Rysanek wrote:
>> >Dear gentlemen,
>> >
>> >I've just tried looking at the CPU temperature via coretemp.ko on
>> >some SandyBridge CPU's. I have a desktop-grade Core I3 at 3.3 GHz,
>> >and a mobile Core i5 at 2.5 GHz. When I flip the CPU from "pretty
>> >much idle" to "each core running a cpuburn instance", the temperature
>> >reported by coretemp*jumps*  up in a second. On the desktop CPU, the
>> >jump is about 10*C. On the mobile CPU, the jump is about 25-30 *C !
>> >It jumps up in a second after I launch the software load - and then
>> >it proceeds to gradually inch up, as the massive heatsink starts to
>> >warm up a bit (perhaps another degree C in twenty seconds to a
>> >minute).
>> >
> I think that is pretty normal. Keep in mind this is the temperature within
> the chip core, not at the external package.
>
> I have systems with i7-2600 and i7-3770K. Both run at 30C or less when idle.
> The core temperature jumps by at least 10C as soon as I add load. If the
> load is at 100% for a minute or so on all cores, the temperature increases
> to around 80C, and the CPU fan speed increases to keep it there.
> Nothing to be concerned about unless your fan control is not working.
>
>> >Should I look for some thermal mischief in my system?
>> >Or, do the SandyBridge CPU's indeed "precompensate" the temperature
>> >reading based on instantaneous power consumption?
> Not really. It is just that with higher power consumption, the CPU temperature
> can jump significantly within a few seconds. After all, the power translates
> into produced heat, which has to go somewhere. Since the sensors are directly
> at the core, you see the impact immediately.
>
> I think that may be a bit different to older CPUs, which probably tend to run a
> bit hotter even when idle, resulting in less temperature increase under load.
> Or, in other words, Intel probably did a good job conserving power when it
> is not needed with recent CPUs.
>
>> >I.e., is the "digital thermal sensor" more like a "fan control hint
>> >with a strong feed-forward component", rather than a half-decent
>> >thermometer?
>> >
>> >I've noticed already on some recent 45nm Core2 CPU's (on a gigabyte
>> >motherboard) that the fan control feedback loop responds magically
>> >swiftly to me launching some software load.
>> >I know that modern SuperIO chips (containing PWM fan control logic)
>> >can take the CPU temp value straight from the CPU (perhaps mediated
>> >by the south bridge), using a digital link called PECI. I also know a
>> >bit about the autonomous fan control algorithms implemented in chips
>> >by ITE and Winbond(Nuvoton) - I've played with these before.
>> >If indeed the CPU's DTHERM sensor would "bump up the reading" in
>> >response to power consumption, that would explain the swift
>> >modulation of fan speed based on instantaneous CPU load...
>> >
> power -> heat -> increased temperature ... no need to be too fancy.
>
> Hope this helps,
>
> Guenter
>


_______________________________________________
lm-sensors mailing list
lm-sensors@lm-sensors.org
http://lists.lm-sensors.org/mailman/listinfo/lm-sensors

Re: [lm-sensors] coretemp.ko -- Sandy Bridge: temperature readings "biased" based on power consumpti

[Frantisek Rysanek]

On 31 Oct 2012 at 6:31, Phil Pokorny wrote:
> 
> On Tue, Oct 30, 2012 at 5:32 PM, Guenter Roeck <linux@roeck-us.net> wrote:
[...]
>    > I think that is pretty normal. Keep in mind this is the 
>    > temperature within the chip core, not at the external package.
> 
> And there is internal thermal paste inside the CPU package between the
> CPU die itself and the integrated heat spreader. That thermal material
> also explains why the silicon die temperature can jump so quickly. It
> has a higher deg C/W than the IHS and the heatsink on top of that.
> 
> Phil P.

Hmm. Okay - I should check my thermocoupling again.

The CPU that "jumps up 30*C in a second" (and jumps down
equally fast when I turn the load off) is actually a mobile/embedded 
version, i.e. it has no I.H.S. The chip is supposedly thermocoupled 
straight flat to the heatsink. This is a passive heatsink, and it's 
pretty massive, and I know that heatsinks tend to have quite some 
thermal inertia. Theoretically this should be the optimum style of 
thermocoupling in "passive cooled" computers.

The catch is, that this particular hardware setup had flawed 
thermocoupling between the CPU and the heatsink (ex works). The CPU 
had a pad of "thermocouple chewing gum" on it (likely Fuji Sarcon), 
approx. 1.1 mm thick. That way the CPU reported 80*C in the BIOS 
Setup a few dozen seconds after power-up :-)  = one core running in a 
tight loop, other cores asleep, heatsink at 20*C, with the whole 
computer consuming roughly 27 Watts. (Under Linux, when properly 
idling, consumption dropped to 12 Watts.)

I've tried to fix the problem by inserting a sheet of 1mm copper 
(about 4 by 4 cm) and by glueing it all toghether by Ceramique CMQ2 - 
I couldn't use the Coollaboratory liquid (Ga+In+Sn), because the 
heatsink is made of Aluminium... In a similar scenario, I used to get 
maybe 10*C of a temperature gradient between a Core2 Duo and the 
heatsink, and that was across a rather long aluminium "thermocouple 
bridge".
In the case at hand I was fairly satisfied with the way the CPU leans 
against my improvised heat spreader - I could feel it by the 
screwdriver, as I was tightening the mounting bolts. Also, the square 
of the SandyBridge chip is *big* compared to C2D Mobile = more heat 
transfer surface for a similar wattage. Unfortunately there's no way 
for me to inspect the gap with the heatsink mounted, so I'll probably 
have to trust the thermal sensor :-)

I still plan to keep an eye on the mobile SandyBridge CPU's whenever 
I get a chance, to get some "benchmark" / comparable readings on 
other hardware. I don't meet too many SandyBridge CPU's in our 
embedded computers - the boxes are dominated by the Atom and I have 
to agree that even mobile SandyBridge is thermally a bit of an 
overkill in those small boxes (just like the mobile C2D was). 
Besides, SandyBridge is stil a bit of a newcomer in the 
industrial/embedded hardware. There are some SandyBridge notebooks 
around - so I'll probably have to ask my boss to let me PXE-boot into 
Linux to run a comparison test on his gear...

I can see a similar jump (20-25*C) in a Core i5 Arrandale -based 
notebook, where I have reasons to believe that the thermocoupling is 
done right.

Hmm... interesting stuff :-)
Thanks for your insight - to everybody who responded.

Frank Rysanek


_______________________________________________
lm-sensors mailing list
lm-sensors@lm-sensors.org
http://lists.lm-sensors.org/mailman/listinfo/lm-sensors