PiCAP A Parallel and Incremental Capacitance ExtractionPICAP并行和增量电容提取.pptVIP

* * To handle the problem, there are many works on stochastic capacitance extraction. For example, stochastic integral by low-rank approximate method is proposed, but it is slow for full-chip extraction. Another is the stochastic orthogonal polynomial expansion method, which results in an augmented and dense system Therefore, our objective is mainly to improve the efficiency of variational capacitance extraction. First, FMM with nearly linear complexity can be used to reduce time for MVP, and further be parallelized. Also, for different variation, preconditioner should be updated incrementally. * * Now, let’s introduce the stochastic geometric moment first. * In order to do so, we develop a stochastic capacitance extraction method, called piCAP. The entire computing flow is shown as this figure. First, piCAP makes use of geometric moments to integrate process variation into potetial coefficient matrix. In other words, Pij can be expressed with an explicit geometry dependence, which can capture process variation correctly. Thus, we develop a modified FMM so that we can evaluate MVP p*q with variation in parallel. Furthermore, we develop iGMRES to solve the variational system. We use spectral preconditioner to improve iteration convergence, and we incrementally update it for different variation. * * * In this paper, we consider two stochastic variation sources: panel distance ‘d’ and panel width ‘h’. To define geometric moments, we need first consider multipole expansion along x-y-z coordinates. Assume there are two cubes: source cube and observer cube as shown in this figure. ‘r’, ‘d’ are vectors. We can define the multipole expansion of one over r minus d, which is the potential on center of observer cube due to the unit charge on source panel in the free space. For local expansion, denotes by Li, can be simply achieved by exchanging d and h in Mi.Clearly, both Mi and Li have an explicit dependence on panel width h and distance d. We call them geometric moments.