Details of the abstract
Title of paper | Hybrid GPU solution to regularized divergence-free curl-curl equations for electromagnetic inversion problems |
List of authors | Dong, H., Sun, K., Egbert, G., Kelbert, A., Meqbel, N. |
Affiliation(s) |
China University of Geosciences (Beijing), Nvidia Corporation, Oregon State University, United States Geological Survey, 3D Consulting-GEO GmbH. |
Summary |
The Curl-Curl equation is the key to the time-harmonic electromagnetic (EM) problems in geophysics. The efficiency of its solution is decisive to the performance of EM simulations, which account for over 90% of the computation cost in inversions like Magnetotellurics or controlled source EM problems. However, most published EM computation codes are still CPU-based and cannot utilize the recent computation techniques with GPUs. Based on the previously proposed divergence-free algorithm developed on CPUs, this study demonstrates the current limits of the CPU-based inversion procedure. To exploit the high throughput computational ability of GPUs, the study proposes a hybrid CPU-GPU framework to solve the forward and adjoint problems of the EM inversions. The large sparse linear systems arise from the staggered-grid finite difference approximation of curl-curl problems are solved with a new mixed-precision Krylov subspace solver. The algorithm is implemented with the ModEM modular inversion package and with both synthetic and real-world magnetotelluric examples. The results show a promising 15-30x speed-up for the solution stage of the curl-curl equations over single-CPU calculations. On real-world inversion test cases, the overall performance of a GPU-attached computation node with the new hybrid framework is comparable to that of four CPU-only nodes with conventional ModEM implementation. This would make the large-scale frequency domain EM inversions possible on smaller modern GPU platforms with reduced carbon footprint. |
Session Keyword | 2.0 Theory, Modelling and Inversion |
File upload |
2.0_hybrid_gpu_solution_to_re_dong_01.pdf
|