Skip to article frontmatterSkip to article content
Simula Research Laboratory

Leveraging DOLFINx data-oriented design for GPU implementation

Authors
Affiliations
University of Cambridge
University of Cambridge
University of Cambridge

The data-oriented design choice adopted by DOLFINx offers several advantages. One such benefit is the ease of implementing kernels that can run on both CPU and GPU architectures. In this presentation, we discuss our efforts in implementing the mass and stiffness kernel on the GPU. Utilizing the Numba CUDA interface, we demonstrate how the data-oriented design simplifies the implementation process for GPU acceleration. Furthermore, we present performance results to illustrate the efficiency of our implementation, despite its simplicity. Specifically, using a benchmark problem in transcranial ultrasound simulation, we showcase comparable performance achieved on a modest GPU architecture compared to a single node on an HPC architecture.

Abstracts
Integrating Mechanistic Knowledge into Deep Learning for Improved Cancer Detection
Abstracts
Magnetic Radiation and Thermoelectric Coupling in a two-variable model of the cardiac action potential: Effects of excessive use of Gadgets