Copyright by William Lloyd Bircher 2010 - The Laboratory for ...
Copyright by William Lloyd Bircher 2010 - The Laboratory for ...
Copyright by William Lloyd Bircher 2010 - The Laboratory for ...
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
Temperature is controlled <strong>by</strong> adjusting the speed of the processor’s fan. Temperature is<br />
observed with a 1 /8 degree Celsius resolution using an on-die temperature sensor [Bk09].<br />
This sensor can be accessed <strong>by</strong> the system under test through a built-in, on-chip register.<br />
<strong>The</strong> resultant temperature-dependent leakage equation is shown in Table 3.9. Since<br />
temperature is modeled over only the operating range of the processor, it can be<br />
accounted <strong>for</strong> as a quadratic equation. Alternatively, a wider temperature range can be<br />
accounted <strong>for</strong> using and an exponential equation in the <strong>for</strong>m of a×e T×b . <strong>The</strong> coefficients a<br />
and b are found through regression. <strong>The</strong> term T represents the die temperature in Celsius.<br />
For this study the quadratic <strong>for</strong>m is used due to its lower computational overhead and<br />
sufficient accuracy. Voltage is controlled using the P-State Control Register [Bk09].<br />
This allows selection of one of five available voltage/frequency combinations. Voltage is<br />
observed externally as a subset of the traced power data. Like the workload dependent<br />
model, the coefficients of the static power model are tuned using regression techniques.<br />
Note that the static power model is highly process dependent. Processors with different<br />
semiconductor process parameters require the model to be re-tuned.<br />
<strong>The</strong> dominant power management effects (voltage/frequency scaling, clock gating) are<br />
further accounted <strong>for</strong> using the gateable and ungateable power models. Gateable power<br />
is found <strong>by</strong> measuring the effect of enabling/disabling idle core clock gating. Ungateable<br />
represents the portion of power which cannot be gated. <strong>The</strong>se components are also found<br />
experimentally. <strong>The</strong> resultant, average error in the model is 0.89%. <strong>The</strong> standard<br />
deviation of the error <strong>for</strong> SPEC CPU2006 and SYSmark 2007 is less than 1%. Worst-<br />
38