Accelerating cardiac bidomain simulations using graphics processing units.

Area of Science:

• Computational biology
• Cardiac electrophysiology
• High-performance computing

Background:

• Multiscale computer models of the heart are crucial for understanding cardiac function and disease.
• Computational demands of these detailed simulations limit their widespread use.
• High-performance computing (HPC) and acceleration technologies like GPUs offer potential solutions.

Purpose of the Study:

• To demonstrate the feasibility and performance of multi-GPU bidomain simulations.
• To assess the scalability of GPU acceleration for cardiac modeling.
• To evaluate the efficiency of GPUs compared to CPUs for solving large sparse linear systems in bidomain models.

Main Methods:

• Utilized a state-of-the-art finite element method (FEM) model of rabbit ventricles on unstructured grids.
• Adapted existing Cardiac Arrhythmia Research Package (CARP) code for multi-GPU execution with minimal changes.
• Conducted strong scalability benchmarks on 6-20 GPUs.

Main Results:

• Achieved simulation speedups of 11.8 to 16.3 times on GPUs compared to an equivalent number of CPU cores.
• A simulation utilizing 20 GPUs required only 476 CPU cores for comparable performance.
• Demonstrated the effectiveness of GPUs for accelerating bidomain simulations.

Conclusions:

• Multi-GPU implementation is feasible and highly effective for accelerating cardiac bidomain simulations.
• GPU acceleration offers a significant advantage over traditional CPU-based HPC for complex cardiac modeling.
• This advancement can facilitate wider adoption of in-silico cardiac research.