News:

Follow BoincTasks on Twitter Facebook        Visit our website here.
BoincTasks cloud login is working again

Main Menu

by-Core offsets

Started by archae86, June 08, 2012, 10:22:21 PM

Previous topic - Next topic

0 Members and 1 Guest are viewing this topic.

archae86

I wish TThrottle would support user-specified offsets to the reported temperature by core.  This would allow much more consistent throttling for those of us running less than 100% of cores on BOINC.

Background: The on-die temperature sensors used on Intel microprocessors were originally put there to support a fleet-wide reliability objective by reducing the amount of excess-temperature abuse seen in service.  So far as I know, they, and the control systems in which there are embedded, are serving that function adequately, but there are actually rather bad thermometers.  Surprisingly enough, they are often not well-matched from core to core on the same die.  While users often think such mismatches are symptoms of poor application of thermal grease, misaligned heat sinks, and such, offsets are near zero power are not from those causes, and often offsets at high power are not either.

For TThrottle users who run BOINC applications on every available core, all of this is not an issue, as they can just dial in a desired boundary suitable for the highest-reading core, and all else tracks.

But on my GPU-equipped machines, I've decided, as do many others, to run less than a full set of CPU applications (most just leave one idle core--whereas I am running only two CPU tasks on one four-core non-hyperthreaded Sandy Bridge and also on one Nehalem with eight virtual cores).  Given the large temperature offset of the sensors on my cores, this means the real amount of throttling varies just because of which core Windows happens to assign my running BOINC CPU tasks to.  On Windows 7 these assignments seem to be stable over many minutes of time--without the hectic shifting seen under Windows XP.

Speedfan, for a well-known example, supports per-core user-specified temperature offsets in the form of an integer to be added or subtracted for each core from the reported reading.  While this can't handle some of the more obscure errors, including missing codes, slope error, and such, it is good enough for most purposes for most of the CPUs I've owned or heard reports on, and I think would be a nice addition to Throttle.

fred

Quote from: archae86 on June 08, 2012, 10:22:21 PM
I wish TThrottle would support user-specified offsets to the reported temperature by core.
I will add this to the tthrottle.xml file

<PROCESSOR>
  <Core_adjust>1;-1;0;-2</Core_adjust>
</PROCESSOR>

archae86

Quote from: fred on June 09, 2012, 02:26:19 PMI will add this to the tthrottle.xml file

<PROCESSOR>
  <Core_adjust>1;-1;0;-2</Core_adjust>
</PROCESSOR>
Wonderful.  As not all applications which reference Intel Cores number them the same, I'd assume users can figure what will work for them in yours by checking the temperatures shown in the temperature tab of TThrottle.

By the way, I am a former Intel person, and in the distant past was a microprocessor design person, then reliability, then factory data analysis for yield improvement.  I won't, precisely, claim inside knowledge for my comments on sensor historical purpose and sensor error, but am actually pretty confident in them.