Re: [gentoo-user] Re: lm_sensors much hotter in 2.6.31 than 2.6.28

[Paul Hartman]

On Mon, Oct 26, 2009 at 2:44 PM, Grant <emailgrant@×××××.com> wrote:
>>> After upgrading from 2.6.28 to 2.6.31, I noticed my CPU temperatures
>>> are reported a full 20C hotter.  If I load the old kernel, the
>>> reported temperatures drops back down to normal.  Has anyone else seen
>>> this?
>>
>> If you're using coretemp as sensor, the temps are always off (the coretemp
>> sensor of Intel chips is not accurate, not by any stretch of the
>> imagination.)  It only reports the distance to the CPU's maximum thermal
>> junction, which then the coretemp driver *tries* to translate into a
>> temperature, but the result is wrong since the value reported by the CPU is
>> not accurate to start with (it only gets accurate as you approach the max
>> value).  That maximum value is totally undocumented for desktop CPUs (the
>> docs Intel provided recently are wrong.)
>>
>> You should use your mainboard's sensors instead for accurate values.
>
> I'm actually using k8temp.  Do you think it is susceptible to the same
> problems you're talking about?  I also have an ACPI sensor available
> named THRM.  Should that one be more accurate?
>
> BTW, another system of mine (Dell laptop) only seems to have available
> coretemp or an ACPI sensor which reports values like 46960 mWh.  Am I
> totally out of luck with that one?
>
> So, In the end, it's fairly impossible to monitor a CPU's actual
> temperature in order to keep it below the published maximum?

Hmm, the k8temp documentation seems to indicate that it should be
actual temperature:

"Temperatures are measured in degrees Celsius and measurement
resolution is 1 degree C. It is expected that future CPU will have
better resolution. The temperature is updated once a second. Valid
temperatures are from -49 to 206 degrees C."

Also, with lm_sensors not all sensors can be auto-detected. I had to
manually specify mine (Abit uGuru3).