Abstract Details

Details of the Abstract

Title of paper Multi-GPU accelerated parallel modelling and inversion for geophysical electromagnetic problems
List of authors Hao Dong, Kai Sun, Gary Egbert, Anna Kelbert, Naser Meqbel
Affiliation(s) China University of Geosciences, Beijing, NVIDIA Corporation, Beijing, Oregon State University, Corvallis, OR, United States Geological Survey, Golden, CO, 3D Consulting-GEO GmbH, Berlin
Summary The Curl-Curl equations is the fundament for time-harmonic electromagnetic (EM) problems in geophysics. The forward and adjoint calculations with the discretized Curl-Curl equations are computationally intensive and crucial to frequency domain geophysical EM inversions like Magnetotellurics and Controlled Source EM. However, most published algorithms are still CPU-based and cannot benefit from the rapid development of modern large-scale multi-GPU parallel architectures driven by the AI boom. Based on the previously proposed divergence-free modelling algorithm for CPUs, we develop a hybrid parallel paradigm to exploit the high-throughput of interconnected heterogenous parallel systems equipped with multiple GPUs. The large sparse linear system derived from the staggered-grid finite-difference approximation of Curl-Curl problem is decomposed into sub-domains to mitigate the limitation of the memory size of a single GPU card. The system is then solved efficiently with a mixed-precision Krylov subspace GPU algorithm in parallel, utilizing the fast device-to-device interconnection of NVIDIA NVlink bridges. To show how the large-scale inversion problems can be substantially accelerated, we implement the new algorithm in the ModEM framework and test the algorithm with real-world 3D problems. We examine the efficiency of our new algorithm in forward and adjoint calculations with an MT problem with ~4 million model cells and 723 stations. The results from a GPU node equipped with 6 consumer-grade NVIDIA RTX 3090 cards demonstrate a promising ~2.5x speed-up, when compared with conventional CPU-based algorithm on a server cluster with 12 nodes. This development may significantly reduce the computation time and carbon footage for large-scale frequency domain EM inversion problems and brings the possibility of near real-time EM imaging, for engineering and environmental applications.
Session Keyword 2.0 EM theory, modelling and Inversion
File upload 2.0_multi-gpu_accelerated_par_dong.pdf
Back to list