Available at: https://digitalcommons.calpoly.edu/theses/3318
Date of Award
6-2026
Degree Name
MS in Electrical Engineering
Department/Program
Electrical Engineering
College
College of Engineering
Advisor
Andrew Danowitz
Advisor Department
Electrical Engineering
Advisor College
College of Engineering
Abstract
The Finite-Difference Time-Domain (FDTD) method is a numerical technique for solving partial differential equations. It was first developed to solve Maxwell’s equations for electromagnetic wave propagation and has since been extended to model other physical phenomena governed by wave propagation. Because the FDTD method is data-parallel, it is well-suited for GPU acceleration; modern FDTD-based simulations run offline on large GPU clusters. There is, however, very little research on running FDTD on resource-constrained embedded GPUs, which are increasingly popular for real-time applications.
This thesis explores a CUDA-based FDTD solver for the 3D wave equation on the Nvidia Jetson Orin Nano. This work evaluates the trade-offs between speed, accuracy, and energy efficiency across three implementation categories: thread allocation strategy, memory management approach, and numerical precision format. The results show that the choice of precision format is the most important factor: Floating point 32 (FP32) is nearly as accurate as floating point 64 (FP64) while taking about half the time and energy for a given computation. Floating point 16 (FP16) is the fastest and most energy-efficient precision choice, but up to 1000 times less accurate. Fixed-point integer formats draw less instantaneous power but run much longer, consuming more total energy, and are tens of thousands of times less accurate than FP64. Explicit memory copies between the CPU and GPU are up to 75% faster than using managed memory, even though the underlying hardware features unified memory between the CPU and GPU. The thread allocation strategy has no obvious effect since the GPU is fully occupied in all cases. From these results, we propose guidelines for running FDTD efficiently on embedded GPUs.