Adaptive least mean square behavioral power modeling

Abstract

In this work we propose an effective solution to the main challenges of behavioral power modeling: the generation of models for the power dissipation of technology-independent soft macros and the strong dependence of power from input pattern statistics. Our methodology is based on a fast characterization performed by simulating the gate-level implementation of instances of soft macros within the behavioral description of the complete design. Once characterization has been completed, the backannotated behavioral model replaces the gate-level representation, thus allowing fast but accurate power estimates in a fully behavioral context. Our power characterization procedure is a very efficient process that can be easily embedded in synthesis-based design flows. No additional effort is required from the designer, since power characterization merges seamlessly with a natural top-down design methodology with iterative improvement. After characterization, the behavioral power simulation produces accurate average and instantaneous pourer estimates (with errors around 7% and 25%, respectively, from accurate gate-level power simulation).