Acoustic Full-Waveform Inversion in Julia on Multi-xPUs
DescriptionFull-waveform inversion (FWI) is a high-resolution technique used to infer physical properties of the interior of a medium by running numerical simulations and comparing the results with observed waveforms. The computational cost and complexity of FWI currently make it unappealing to non experts and for applications where time-to-solution is critical. Our work aims to combine close-to-ideal performance and scalability with ease of use and low HPC knowledge requirements to make FWI more accessible to a larger scientific community. We introduce a package for acoustic forward and gradient computations targeting FWI fully written in the Julia programming language that can run on multiple CPUs or GPUs. We show how our finite-difference solver can reach close to peak performance as well as almost linear weak scaling on the GPU nodes of the Piz Daint supercomputer. Its flexibility, efficiency, and accessibility make it a suitable choice for both “rapid prototyping” scenarios as well as complex large-scale inversions.
TimeWednesday, June 2815:00 - 15:30 CEST
Computer Science, Machine Learning, and Applied Mathematics