Stochastic dominant singular vectors method for variation-aware extraction

In this paper we present an efficient algorithm for variation-aware interconnect extraction. The problem we are addressing can be formulated mathematically as the solution of linear systems with matrix coefficients that are dependent on a set of random variables. Our algorithm is based on representing the solution vector as a summation of terms. Each term is a product of an unknown vector in the deterministic space and an unknown direction in the stochastic space. We then formulate a simple nonlinear optimization problem which uncovers sequentially the most relevant directions in the combined deterministic-stochastic space. The complexity of our algorithm scales with the sum (rather than the product) of the sizes of the deterministic and stochastic spaces, hence it is orders of magnitude more efficient than many of the available state of the art techniques. Finally, we validate our algorithm on a variety of on-chip and off-chip capacitance and inductance extraction problems, ranging from moderate to very large size, not feasible using any of the available state of the art techniques.