Parallel Sparse Direct Solver For Integrated Circuit Simulation PDF Download

Abstract : Integrated circuit technology has gone through several decades of aggressive scaling.It is increasingly challenging to analyze growing design complexity. Post-layout SPICE simulation can be computationally prohibitive due to the huge amount of parasitic elements, which can easily boost the computation and memory cost. As the decrease in device size, the circuits become more vulnerable to process variations. Designers need to statistically simulate the probability that a circuit does not meet the performance metric, which requires millions times of simulations to capture rare failure events. Recent, multiprocessors with heterogeneous architecture have emerged as mainstream computing platforms. The heterogeneous computing platform can achieve highthroughput energy efficient computing. However, the application of such platform is not trivial and needs to reinvent existing algorithms to fully utilize the computing resources. This dissertation presents several new algorithms to address those aforementioned two significant and challenging issues on the heterogeneous platform. Harmonic Balance (HB) analysis is essential for efficient verification of large postlayout RF and microwave integrated circuits (ICs). However, existing methods either suffer from excessively long simulation time and prohibitively large memory consumption or exhibit poor stability. This dissertation introduces a novel transient-simulation guided graph sparsification technique, as well as an efficient runtime performance modeling approach tailored for heterogeneous manycore CPU-GPU computing system to build nearly-optimal subgraph preconditioners that can lead to minimum HB simulation runtime. Additionally, we propose a novel heterogeneous parallel sparse block matrix algorithm by taking advantages of the structure of HB Jacobian matrices as well as GPU's streaming multiprocessors to achieve optimal workload balancing during the preconditioning phase of HB analysis. We also show how the proposed preconditioned iterative algorithm can efficiently adapt to heterogeneous computing systems with different CPU and GPU computing capabilities. Extensive experimental results show that our HB solver can achieve up to 20X speedups and 5X memory reduction when compared with the state-of-the-art direct solver highly optimized for twelve-core CPUs. In nowadays variation-aware IC designs, cell characterizations and SRAM memory yield analysis require many thousands or even millions of repeated SPICE simulations for relatively small nonlinear circuits. In this dissertation, for the first time, we present a massively parallel SPICE simulator on GPU, TinySPICE, for efficiently analyzing small nonlinear circuits. TinySPICE integrates a highly-optimized shared-memory based matrix solver and fast parametric three-dimensional (3D) LUTs based device evaluation method. A novel circuit clustering method is also proposed to improve the stability and efficiency of the matrix solver. Compared with CPU-based SPICE simulator, TinySPICE achieves up to 264X speedups for parametric SRAM yield analysis without loss of accuracy.