Making Power-Efficient Data Value Predictions

摘要

Power dissipation due to value prediction is being more studied recently. In this paper, a new cost effective data value predictor based on a linear function is introduced. Without the complex two-level structure, the new predictor can still make correct predictions on some patterns that can only be done by the context-based data value predictors. Simulation results show that the new predictor works well with most value predictable instructions. Energy and performance impacts of storing partial tag and common sub-data values in the value predictor are studied. The two methods are found to be good ways to build better cost-performance value predictors. With about 5K bytes, the new data value predictor can obtain 16.5% maximal while 4.6% average performance improvements with the SPEC INT2000 benchmarks.