dGPLVM: A Nonparametric Device Model for Statistical Circuit Simulation

In this study, a novel statistical device model, dGPLVM, is proposed. In dGPLVM, device variation is compactly represented using low-dimensional latent variables and temperature dependence can be incorporated. The fitting accuracy of the dGPLVM and the generated device characteristics obtained through sampling in the latent space are validated using the measurement results of commercial power MOSFETs. Verilog-A implementation of the proposed method demonstrates its practicality.