A multi-GPU benchmark for 2D Marchenko Imaging

TL;DR

This paper introduces a multi-GPU implementation of an iterative Marchenko imaging algorithm, significantly accelerating seismic Green's function estimation from surface reflection data compared to CPU-based methods.

Contribution

The work presents a GPU-accelerated version of the Marchenko method using segmented dot products and CUDA, enabling faster processing of large seismic datasets.

Findings

GPU implementation outperforms CPU version in speed

Efficient acceleration of convolution and integration steps

Applicable to large-scale seismic reflection data

Abstract

The Marchenko method allows estimating Green's functions with a virtual source in the subsurface from a reflection response on the surface. It is an inverse problem that can be solved directly or by an iterative scheme, with the latter being more feasible computationally. In this work we present a multi-GPU implementation of a well-established iterative Marchenko algorithm based on (the) Neumann series. The time convolution and space integration performed on each iteration, also referred to as synthesis, are here represented as a segmented dot product, which can be accelerated on modern GPUs through the usage of warp-shuffle instructions and CUDA libraries. The original CPU version is benchmarked on 36 CPU cores versus the implemented version on 4 GPUs, over three different reflection data sets, with sizes ranging from 3 GB to 250 GB.