Fast Analog Circuit Synthesis Using Sensitivity Based Near Neighbor Searches

We present an efficient analog synthesis algorithm employing regression models of circuit matrices. Circuit matrix models achieve accurate and speedy synthesis of analog circuits. In this paper, synthesis is accelerated by eliminating numerous computations of the matrix elements during a synthesis run. Computations are avoided by reusing exact or nearby design points visited during previous synthesis iterations. Hashing and multidimensional nearest neighbor lookup are used in incremental evaluation of design solutions encountered during synthesis. Sensitivity of the design variables is considered for locating a neighboring solution. Neighbor lookup is efficiently performed using box-decomposition trees. The proposed method is used to synthesize three benchmark circuits. Results show that with hashing and neighbor lookup, synthesis is 6x-13x faster than with the use of matrix models alone.