STYLE: a statistical design approach based on nonparametric performance macromodeling

摘要

Despite their potential benefits, statistical design/optimization techniques are often not routinely used by industry because of prohibitive costs and/or lack of generality. To overcome these obstacles, we present a general approach to statistical design which is computationally inexpensive and applicable to an extensive spectrum of analog-digital circuits, unlike previous inexpensive methods which limit the user to specific technologies, standard device models, and/or purely digital designs. This general applicability and cost efficiency is maintained through four factors. First, an efficient screening algorithm minimizes the dimensionality of the input parameter space by identifying the important input parameters. Second, a general, accurate, parsimonious nonparametric regression model is formulated so that a wide range of behavior between input parameters and output performances can be mapped, not just those with a linear regression relationship. Third, this approach tightly couples with a compiler-methodology based simulator which sustains the achieved flexibility through its model independence. Finally, a waveform analysis capability completely automates the procedure. An industry ECL bandgap regulator circuit illustrates the benefits of this method.