Efficient parameter-extraction of SPICE compact model through automatic differentiation

Abstract

A novel parameter extraction method for compact MOSFET models is proposed. The proposed method exploits automatic differentiation (AD) technique that is widely used in the training of artificial neural networks. In the AD technique, gradient of all the parameters of the MOSFET model is analytically calculated as a graph to reduce computational cost. On the basis of the calculated gradient, the model parameters are efficiently optimized. Through experiments using SPICE models, the parameter extraction using the proposed method achieved 7.01x speedup compared to that using the numerical-differentiation method.