Multi-GPU accelerated three-dimensional FDTD method for electromagnetic simulation.

Area of Science:

• Biomedical Engineering
• Computational Science
• High-Performance Computing

Background:

• Numerical simulations using the finite-difference time domain (FDTD) method are vital in biomedical engineering.
• Current FDTD methods with numerical human models face challenges in calculation speed and large-scale computing.
• Optimizing computational efficiency is essential for advancing complex biomedical simulations.

Purpose of the Study:

• To enhance the calculation speed of three-dimensional FDTD simulations for numerical human models.
• To adapt existing FDTD code for a multi-GPU environment utilizing Compute Unified Device Architecture (CUDA).
• To evaluate the performance of a multi-GPU setup against single-GPU and supercomputer configurations.

Main Methods:

• Adaptation of three-dimensional FDTD code to a multi-GPU environment using CUDA.
• Implementation utilizing NVIDIA Tesla C2070 GPGPU boards.
• Performance evaluation comparing multi-GPU, single-GPU, and vector supercomputer systems.

Main Results:

• A four-GPU system achieved approximately 3.5 times the calculation speed of a single GPU.
• The multi-GPU system was slightly slower (approx. 1.3 times) than a vector supercomputer.
• Significant improvements in calculation speed were observed with an increasing number of GPUs.

Conclusions:

• Adapting FDTD code to a multi-GPU environment offers substantial speed improvements for numerical human model simulations.
• Multi-GPU computing presents a viable and effective strategy for accelerating large-scale biomedical simulations.
• Further expansion of GPU numbers promises continued enhancements in computational performance for FDTD methods.