high performance computing on graphics processing units: hgpu.org

Tranmural electrophysiological imaging (TEPI) is becoming a possibility with the aid of 3D in silico cardiac EP models and the statistical estimation theory. By quasi Monte-Carlo (MC) simulation of the 3D EP models on the subject-specific anatomical model, complex and physiologically meaningful spatiotemporal priors are produced to achieve the 2D-to-3D transition of EP data, an inverse problem that otherwise has no unique solutions. While it is acknowledged that macroscopic phenomenological models are more suited for the purpose of inverse problems, the nature of these algorithms still incurs tremendous computational cost that hinders their clinical translation, particularly caused by the MC simulation of high-dimensional, nonlinear EP models and large-matrix operations involved in probabilistic estimation. In this paper, we explore the use of Graphic Processing Units (GPU) to accelerate TEPI because of its high parallelism and large bandwidth. While initial steps have been taken towards GPU acceleration of whole-heart EP simulation using complex, ionic models, few effort is reported on GPU acceleration of subjectspecific, data-driven EP imaging. In this study we will show how we take advantage of the high level of parallelism available in the hardware resources in GPUs and achieve a preliminary but important 16x speedups compared to the most high-end CPU version of TEPI. We also present benchmarking on 3 different GPU devices, point out the bottlenecks that limit the performance, and give guidelines on balancing the cost versus performance tradeoffs for clinical and researcher environments.