| The finite-difference time-domain (FDTD) technique is a numerical method for the solution of electromagnetic field problems. FDTD has a large numerical but a low analytical expense. The FDTD method is easy to implement and has been used successfully for simulating the performance of a broad class of microwave structures. Additionally, it can address problems with inhomogeneous, anisotropic, and nonlinear material properties. This method is ideally suited for a massively parallel hardware implementation due to the inherent parallelism present in the FDTD algorithm. Existing parallel implementations of the FDTD method do not fully exploit this inherent parallelism. Further, some methods do not address the mass memory requirements for large simulations, some suffer from communication bottlenecks, and none can meet the current needs required for performing the state-of-the-art simulations we want to accomplish.; In this thesis, I describe a modular accelerator which uses application-specific integrated circuits to provide massive parallelism for FDTD computation. In contrast to previous approaches, this accelerator architecture maximizes computational parallelism while minimizing the die area. It overcomes the communication bottlenecks to dramatically reduce execution times for large full-wave electromagnetic simulations. This architecture is a scalable, cost-effective implementation of a full three-dimensional FDTD solver. |
