Accelerating three-dimensional FDTD calculations on GPU clusters for electromagnetic field simulation.

Area of Science:

• Biomedical Engineering
• Computational Electromagnetics
• High-Performance Computing

Background:

• Electromagnetic simulations using anatomically realistic computational human models are crucial in biomedical engineering.
• The finite-difference time domain (FDTD) method is widely used for these simulations.
• Improving calculation speed and enabling large-scale computing are key challenges.

Purpose of the Study:

• To adapt three-dimensional FDTD code to a multi-GPU cluster environment for enhanced performance.
• To investigate the computational speed and scalability of the FDTD method on a multi-GPU cluster.
• To enable large-scale electromagnetic simulations with computational human models.

Main Methods:

• Adaptation of 3D FDTD code to a multi-GPU cluster environment utilizing Compute Unified Device Architecture (CUDA) and Message Passing Interface (MPI).
• Implementation on a cluster system with three nodes, each equipped with seven NVIDIA Tesla C2070 GPU boards.
• Performance evaluation of FDTD calculations on the multi-GPU cluster system.

Main Results:

• The multi-GPU cluster system demonstrated significantly faster FDTD calculation speeds compared to a single multi-GPU workstation.
• The GPU cluster system outperformed a vector supercomputer in terms of calculation speed.
• The system's architecture allowed for large-scale FDTD calculations, leveraging over 100 GB of GPU memory.

Conclusions:

• Adapting FDTD code to a multi-GPU cluster environment effectively accelerates electromagnetic simulations with computational human models.
• The developed multi-GPU cluster system offers a powerful and efficient platform for large-scale biomedical engineering simulations.
• This approach represents a significant advancement in computational speed and capability for complex electromagnetic modeling.