Accelerating thermal deposition modeling at terahertz frequencies using GPUs

Abstract

Finite-difference time-domain (FDTD) methods are widely used to model the propagation of electromagnetic radiation in biological tissues. High-performance central processing units (CPUs) can execute FDTD simulations for complex problems using 3-D geometries and heterogeneous tissue material properties. However, when FDTD simulations are employed at terahertz (THz) frequencies excessively long processing times are required to account for finer resolution voxels and larger computational modeling domains. In this study, we developed and tested the performance of 2-D and 3-D FDTD thermal propagation code executed on a graphics processing unit (GPU) device, which was coded using an extension of the C language referred to as CUDA. In order to examine the speedup provided by GPUs, we compared the performance (speed, accuracy) for simulations executed on a GPU (Tesla C2050), a high-performance CPU (Intel Xeon 5504), and supercomputer. Simulations were conducted to model the propagation and thermal deposition of THz radiation in biological materials for several in vitro and in vivo THz exposure scenarios. For both the 2-D and 3-D in vitro simulations, we found that the GPU performed 100 times faster than runs executed on a CPU, and maintained comparable accuracy to that provided by the supercomputer. For the in vivo tissue damage studies, we found that the GPU executed simulations 87x times faster than the CPU. Interestingly, for all exposure duration tested, the CPU, GPU, and supercomputer provided comparable predictions for tissue damage thresholds (ED 50 ). Overall, these results suggest that GPUs can provide performance comparable to a supercomputer and at speeds significantly faster than those possible with a CPU. Therefore, GPUs are an affordable tool for conducting accurate and fast simulations for computationally intensive modeling problems.