GIT-CERCS-03-03
Mongkol Ekpanyapong, Pinar Korkmaz, Hsien-Hsin S. Lee,
Choice Predictor for Free
Reducing energy consumption has become the first priority in
designing microprocessors for all market segments including embedded, mobile,
and high performance processors. The trend of state-of-the-art branch predictor
designs such as a hybrid predictor continues to feature more and larger
prediction tables, thereby exacerbating the energy consumption. In this paper,
we present two novel profile-guided static prediction techniques--- Static
Correlation Choice (SCC) prediction and Static Choice (SC) prediction for
alleviating the energy consumption without compromising performance. Using our
techniques, the hardware choice predictor of a hybrid predictor can be
completely eliminated from the processor and replaced with our off-line
profiling schemes. Our simulation results show an average 40% power reduction
compared to several hybrid predictors. In addition, an average 27% die
area can be saved in the branch predictor hardware for other performance
features.