A Comparative Study of Performance of Eigenvalue Solvers for Parallel Vector Fitting in Multiport Tabulated Data Modeling

Modelling of high-speed modules such as electronic packages and non-uniform transmission lines based on multiport tabulated measured or EM simulated data is becoming increasingly important in modern designs. Vector Fitting (VF) was first introduced as an algorithm for system identification via rational function approximation from tabulated data. Since the algorithm is iterative in nature, minimizing its computational cost and parallel efficiency on mixed CPU and GPU environments is critical in reducing the overall time needed for convergence. One of the expensive steps in these parallel VF approaches is computing the complex eigenvalues of thousands of small, square matrices that result from the All-Splitting method of the parallel VF algorithm. The computational expense of this step tends to vary vastly based on the solver as well as the multi-core CPU architecture used, hence it is useful to the designer to know which solver and platform to use for efficient use of the algorithm. For this purpose, a comparative performance study of the state-of-the-art eigenvalue solvers when using prominent multi-core platforms of AMD and Intel is presented in the context of parallel VF. Results demonstrate that the architecture as well as the type of solver used can significantly impact the efficiency.