A GPU-based calculation using the three-dimensional FDTD method for electromagnetic field analysis.

Area of Science:

• Biomedical Engineering
• Computational Electromagnetics
• High-Performance Computing

Background:

• Numerical simulations using the finite-difference time domain (FDTD) method are increasingly common in biomedical engineering.
• The computational expense of FDTD simulations limits their widespread application.
• There is a need for faster simulation methods to enable more complex biomedical modeling.

Purpose of the Study:

• To investigate the acceleration of three-dimensional (3D) FDTD calculations using general-purpose computing on graphics processing units (GPGPU).
• To evaluate the performance of a GPU implementation against conventional CPUs and supercomputers.
• To assess the impact of calculation domain and thread block size on GPU/CPU speedup.

Main Methods:

• Implementation of the 3D FDTD method on a GPU using Compute Unified Device Architecture (CUDA).
• Utilized an NVIDIA Tesla C1060 as the GPGPU hardware.
• Performance benchmarking against a conventional Central Processing Unit (CPU) and a vector supercomputer.

Main Results:

• GPU-accelerated 3D FDTD calculations demonstrated a significant reduction in run time compared to CPU-based methods.
• The performance gains were substantial even for a native GPU implementation.
• The observed speedup ratio between GPU and CPU varied depending on the specific calculation domain and thread block size.

Conclusions:

• GPGPU computing offers a viable and effective approach to accelerate computationally intensive 3D FDTD simulations in biomedical engineering.
• The use of GPUs can overcome the speed limitations of traditional FDTD methods, enabling more efficient research and development.
• Further optimization of GPU implementation parameters can potentially yield even greater performance improvements.