Sujal Dave, PhD

Sujal Dave, PhD

GPU-Accelerated Cardiovascular Fluid-Structure Interaction in SimVascular

Overview

High-fidelity Fluid-Structure Interaction (FSI) simulations are essential for advancing our understanding of cardiovascular disease and improving clinical surgical planning. However, the immense computational cost of these multi-physics models acts as a major barrier, limiting routine deployment and large-scale patient-specific analyses such as uncertainty quantification (UQ). Traditional CPU-based solvers are often too slow for practical clinical workflows, with the linear solver alone consuming approximately 70% of the total simulation effort.

To address this bottleneck, this project introduces a hybrid CPU-GPU approach leveraging GPU acceleration and advanced preconditioning within the open-source SimVascular ecosystem. By modifying the svMultiPhysics solver, we kept the finite element model (FEM) assembly and nonlinear FSI iterations on the host CPU, while offloading the computationally dominant linear solver phase to the device GPU. This device acceleration handles sparse matrix-vector products, Krylov-based iterative solvers, vector updates, and preconditioner applications using the Trilinos framework (Tpetra and Belos) and Kokkos portable GPU backends. This strategy drastically reduces total wall-time while preserving the existing SimVascular multiphysics architecture, moving cardiovascular biomechanics closer to rapid clinical translation.

Results

  • Shifting from legacy CPU-only executions to GPU configurations, combined with a careful choice of preconditioner settings, yielded up to a 90% total reduction in simulation wall time, particularly at higher processor counts.

  • Under legacy settings using CPUs with diagonal preconditioning, the solver spent up to 98.2% of the total simulation time strictly in the linear solve stage.

  • Implementing Algebraic Multigrid (AMG) preconditioners on advanced NVIDIA Hopper 90 GPUs successfully reduced the linear solver’s computational burden to as low as 16.7%.

  • Ongoing & Future Work: While current generalized preconditioners deliver significant speedups, specialized cardiovascular algorithms can still outperform them in specific scenarios. Ongoing efforts are focused on porting custom, in-house linear solvers, specifically a resistance-based preconditioner and the bi-partition linear solver to the GPU as custom Tpetra operators within Trilinos to merge hemodynamics-specific algorithmic efficiency with the computational throughput of modern architectures like the Hopper 90.