Accelerating simulations of cardiac electrical dynamics through a multi-GPU platform and an optimized data structure.

Area of Science:

• Computational biology
• Biophysics
• High-performance computing

Background:

• Cardiac electrophysiological simulations, especially in 3D, are computationally intensive, requiring billions of differential equations.
• Existing methods face challenges with global memory access bottlenecks on Graphics Processing Units (GPUs) due to large spatial domains.
• Accelerating these simulations is vital for understanding cardiac electrical behavior.

Purpose of the Study:

• To develop an efficient computational strategy for accelerating cardiac electrophysiological simulations on GPUs.
• To address memory access bottlenecks in single-GPU and multi-GPU environments.
• To enable faster and more accessible simulations of 3D cardiac tissue.

Main Methods:

• Proposed a specialized data structure and memory access pattern for the diffusion term computation on single GPUs.
• Optimized for coalescent memory transactions and minimized branch divergence.
• Developed a communication strategy for efficient multi-GPU simulations.

Main Results:

• Achieved a 1.4x speedup compared to standard GPU methods for single-GPU simulations.
• Demonstrated efficient performance on multi-GPU platforms through a designed communication strategy.
• Enabled 3D cardiac tissue simulations to run at speeds approaching real-time (only 4x slower).

Conclusions:

• The proposed computational strategy significantly accelerates cardiac electrophysiological simulations on GPUs.
• The approach effectively mitigates memory access bottlenecks in both single- and multi-GPU settings.
• This advancement makes complex 3D cardiac tissue simulations more feasible and accessible.