GPU Accelerated Nonlinear Electronic Circuits Solver for Transient Simulation of Systems with Large Number of Components

David Černý,Josef Dobeš

doi:10.3390/electronics9111819

David Černý, Josef Dobeš

Open Access

https://doi.org/10.3390/electronics9111819

Copy DOI

Journal: Electronics	Publication Date: Nov 2, 2020
Citations: 2	License type: CC BY 4.0

Affiliation: Czech Technical University in Prague

Abstract

GPU cards have been used for scientific calculations for many years. Despite their ever-increasing performance, there are cases where they may still have problems. This article addresses possible performance and memory issues and their solutions that may occur during GPU calculations of iterative algorithms. Specifically, the article focuses on the optimization of transient simulation of extra-large highly nonlinear time-dependent circuits in SPICE-like electronic circuit simulator core enhanced with NVIDIA/CUDA (Compute Unified Device Architecture) interface and iterative Krylov Subspace methods with emphasis on improved accuracy. The article presents procedures for solving problems that may occur during this integration and negatively affect either the simulation speed or the accuracy of the calculation. Finally, a comparison of the implementation of an iterative calculation procedure with the use of GPU cards, calculation by the direct method and calculation on the CPU only is presented.

Highlights

This article proposes a new implementation of computational core of SPICE-like (Simulation Program with Integrated Circuit Emphasis) simulation program of electrical circuits with CUDA/GPU [1] paralelized BicgStab (Biconjugate Gradient Stabilized) method [2]
Modified core of NgSpice performs standard factorization method using LU factorization and resulting solution is passed via the CUDA interface to various implementation of parallelized BiS method on the graphics card
Iterative algorithms will be able to find the result faster. When methods such as GMRES and BicgStab converge to a result without the use of preconditioners

Summary

Introduction

This article proposes a new implementation of computational core of SPICE-like (Simulation Program with Integrated Circuit Emphasis) simulation program of electrical circuits with CUDA/GPU [1] paralelized BicgStab (Biconjugate Gradient Stabilized) method [2]. This article focuses on extending the standard simulation procedure with an iterative BicgStab method [16] In contrast to these works, this paper publishes a new simulation procedure for simulation of the circuits whose mathematical representation leads to the solution of huge nonlinear time-dependent systems. When simulating electrical circuits, we will almost always have something to do with the solution of a linear system of equations defined by a huge sparse matrix. It is a visualization of a sparse matrix for one time data when solving a transient analysis of the CMOS amplifier circuit. The algorithms cannot be applied to matrices created by an electrical circuit simulator Methods such as GMRES or BicgStab mentioned in the article are much more suitable for this

Simulation

Parallelization

Accuracy

Implementation

Results

Conclusions