This paper develops an aggregation-based algebraic multigrid (AbAMG) method to efficiently analyze the power grids. Different from the conventional algebraic multigrid (AMG) scheme, an innovative constructing method of global inter-grid mapping operator is employed to not only enhance the sparsity of coarse grid operator for reducing the computational complexity but also solve the problem with better convergent rate. The proposed method can solve the circuit with size over two millions in 167.6 CPU seconds (including DC analysis, and transient analysis with 50 time steps), and the maximum error is less than 1%. The significant runtime improvement, over 26X faster than the InductWise [1] and over 1.25X faster than the conventional AMG method, and less memory usage, 40% of the memory usage in [1] are demonstrated.
IntroductionWith the deep sub-micron technology, several features of chips (such as larger number of transistors and lower supply voltages) have made the quality of power delivery network become a key factor of high performance designs [2]. Generally, the power delivery network contains enormous amounts of circuit elements and efficient analyzers are necessary. Thus, general circuit solvers, such as SPICE/HSPICE, by using direct methods are not practicable for the power delivery analysis. In the past years, various efficient methods have been developed for the power distribution network analysis. The preconditioned conjugate gradient method was applied in [3], and the hierarchical methods were developed in [4] [5]. Multigrid-like methods were developed in [6] [7] to map the original problem to a reduced system by using the geometric property of circuit. However, these geometric multigrid techniques [6] [7] are hard to handle the coupling effects of mutual inductances.To deal with the coupling effects, an algebraic multigrid (AMG) [9] based approach was developed in [8]. Generally, the mapping operator of AMG method [9] is determined by each row equation of Ae ≈ 0, where A is the system matrix and e is the error vector of unknown variables. Although this mapping operator doesn't need the geometric information of circuit, its quality strongly counts on the selection of coarse grids and only contains the local information of A. Therefore, it may lose several important error terms, and degrade the convergent rate. To remedy this undesirable behavior, [8] proposed an adaptive coarse grids choosing method. However, its choosing strategy needs to reconstruct the mapping operator at each time step and may boost the CPU time.To solve above problems, we develop a global mapping operator construction procedure based on aggregation AMG methods [10] [11]. The idea of aggregation originates in economics [12], where products in a large scale system are aggregated to become a small system. This procedure can significantly reduce the problem size, and still maintain the accurate representation of overall behaviors. An algebraic partition is performed to the fine grids and the system matrix is partitioned into sev...