High performance multi-platform computing for large-scale image-based finite element modeling of bone.

Area of Science:

• Biomechanics
• Computational Modeling
• Medical Imaging

Background:

• Image-based finite element (FE) modeling estimates bone stiffness and strength non-invasively.
• High-resolution imaging generates large datasets challenging traditional FE solvers.
• Bone-specific mesh-free solvers enhance memory efficiency for bone loading simulations.

Purpose of the Study:

• To benchmark the performance of the FAIM bone-specific, mesh-free solver.
• Evaluate performance across Mac, Linux, and Windows operating systems.
• Assess single- and multi-thread CPU and GPU processing capabilities.

Main Methods:

• Utilized a linear gradient-descent solver with standardized uniaxial loading.
• Employed bone models from micro-CT (µCT) and HR-pQCT scans of the radius and tibia.
• Conducted convergence, speedup, memory, and batch performance tests on CPUs and GPUs across different operating systems.

Main Results:

• Achieved time-per-iteration as low as 0.03s using 3 GPUs on an HR-pQCT radius model (6.45 million DOF).
• Demonstrated strong scaling and parallel efficiencies with GPU and CPU parallel processing (3 GPUs or ≤ 10 CPU threads).
• Obtained low errors (≤0.1%) in force, strain energy density, and Von Mises stress with convergence tolerance of 10^-3 or smaller.

Conclusions:

• FAIM software achieves maximum computational efficiency with convergence tolerance set to 10^-4.
• Using 10 CPU threads or 2 GPUs is sufficient for efficient solution times with µCT and HR-pQCT data.
• Less strict convergence tolerances accelerate solution times but increase outcome measure errors.