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 ...
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
Table 3.4 Instruction Linear Regression Model Comparison<br />
Coefficients<br />
Retired<br />
instructions/cyc<br />
29<br />
Completed<br />
instructions /cyc<br />
α0 α1 α0 α1<br />
36.8 5.28 36.3 5.52<br />
Avg Error 5.45% 4.92%<br />
Coefficient of<br />
Determination<br />
0.679 0.745<br />
Of the µop-based models fetched µops is the most representative metric <strong>for</strong> power<br />
consumption. This suggests that µops that do not update the architected state of the<br />
machine still consume a significant amount of power. For the case of cancelled µops,<br />
this is not surprising since these µops did complete execution but were not retired. So,<br />
they would have traversed nearly the entire processor pipeline consuming a similar power<br />
level as retired µops. More surprising is the effect of fetched µops on the power model.<br />
Fetched µops includes retired and cancelled operations. It also includes the remaining<br />
µops that were cancelled be<strong>for</strong>e completing execution. Since fetched µops provides the<br />
most accurate model, cancelled µops must be consuming a significant amount of power.<br />
<strong>The</strong>se models generate minimum and maximum power values (36W – 47W) similar to<br />
what was found on a Pentium 3 (31W-48W) [Be00] with similar µop/cycle ranges (0 –<br />
2.6). <strong>The</strong> stated average error values are found using the validation set described in<br />
Table 3.1.