Fast multidimensional optimization of analog circuits initiated by monodimensional global Peano explorations
Résumé
A fast design space exploration of analog firm intellectual properties (IP) based on Peano-like paths (piecewise linear and monodimensional) is presented. First, the n-dimensional design space is globally explored following those Peano curves, which are obtained by varying only 1 design variable at a time using a fixed step size. Each variable is taken within a given range. During exploration, the best x-percentile points are retained. After varying globally the n variables, a Nelder–Mead simplex optimization is performed using each of the best points as an initial point. Successive p-variable partitioning of the n -dimensional design space (with p⪯¡n) are applied to adapt the simplex optimization to large dimensions. The proposed exploration technique is combined with a simulation-based hierarchical sizing and biasing methodology to size and bias analog firm IPs. This combined approach has been successfully applied to size and bias a Constant Voltage Reference (CVR) in a 5 V SOI 1um technology. The results illustrate the effectiveness and accuracy of the proposed approach.