Abstract Growing computational demands from seismic imaging applications at bp’s high-performance computing center require recurring infrastructure upgrades to deliver higher performance within fixed power and budget constraints. This article presents a comprehensive hardware-software codesign strategy for migrating a fast Fourier transform (FFT) intensive seismic imaging software framework from traditional central processing unit (CPU) architectures to graphics processing unit (GPU) accelerated platforms. The migration leveraged NVIDIA’s 8-way HGX-H100 nodes featuring fully connected NVLink and NVSwitch to support efficient distributed FFT computation, paired with high-capacity non-volatile memory express (NVMe) solid-state drives (SSDs) enabled by GPUDirect Storage to accelerate node-local input/output (I/O). The legacy Fortran codebase was completely rewritten using a dual-layer design: C++ and Compute Unified Device Architecture (CUDA) for performance-critical kernels to maximize compute and memory efficiency and Python for orchestrating inversion workflows to support rapid research prototyping. Performance validation on production datasets from the Thunder Horse and Herschel fields demonstrated up to 90× runtime speedup and 13× improvement in energy efficiency over CPU platforms, while numerical accuracy was preserved with mean relative errors below 1% for acoustic reverse time migration and full-waveform inversion applications against their CPU references. This successful migration demonstrated the performance and energy efficiency advantages of GPU computing for seismic imaging and provided quantitative justification for bp’s first production-scale GPU cluster.
Zhou et al. (Thu,) studied this question.