Regression Models for Behavioral Power Estimation

Article type: Research Article

Authors: Benini, L.^a | Bogliolo, A.^a | Favalli, M.^b | De Micheli, G.^{a; *}

Affiliations: [a] CSL, Stanford University, Stanford, CA 94305, USA | [b] DEIS, University of Bologna, 40136 Bologna, Italy

Correspondence: [*] Correspondence to: Prof. G. De Micheli.

Abstract: Behavioral power estimation is required to help the designer in making important architectural choices. In this work we propose an accurate and general behavioral power modeling approach especially suited for synthesis-based design ows making use of a library of hard macros implementing behavioral operators. Power dissipation models are pre-characterized and back-annotated in a preliminary step. Accurate information on the power dissipation of the used macros can then be collected during behavioral simulation of the synthesized circuit. Our characterization and modeling methodology is based on the theory of linear regression. Optimal linear power models are obtained with methods of least squares fitting and their generalization to a recursive procedure called tree regression. The regression models can be used for pattern-based dynamic power simulation and for probabilistic static power estimation as well. Our behavioral simulator is integrated within PPP, a multilevel simulation engine for power estimation fully compatible with Verilog XL.

DOI: 10.3233/ICA-1998-5201

Journal: Integrated Computer-Aided Engineering, vol. 5, no. 2, pp. 95-106, 1998