A parallel, adaptive multi-point model order reduction algorithm

This paper describes a model order reduction technique for circuit simulation, based on the parallelization of the well-known multi-point PRIMA algorithm. In order to obtain an optimal accuracy of the reduced-order model in the entire frequency range of interest, the reduced models are computed on different expansion points in correspondence of which the errors, between the transfer functions of the original model and of the actual reduced one, exhibit the largest value, in a recursive way. Moreover, since the computation of the error is a computationally expensive routine, this task is parallelized, assuming that each error value is independent of the others and to work with modern multi-core computers or a cluster of workstations. The numerical results show that the parallelized model order reduction algorithm is able to provide accuracy and speed up with respect to the sequential one, for both dense and sparse data sets.