Enhancing Reverse Time Migration Performance with Half-Precision on GPUs for Production Applications

Abstract Reverse Time Migration (RTM) is a computationally intensive process, often limited by memory bandwidth and storage I/O, particularly in large-scale production environments. In this work, we investigate the use of IEEE FP16 (half-precision floating-point) format to improve RTM performance on modern GPUs. Specifically, we store the wavefield in FP16 format in GPU memory and utilize FP16-based data transfers to CPU memory and disk, significantly reducing memory usage and I/O bandwidth without modifying the core computation in FP32. Synthetic and field data experiments demonstrate that this strategy introduces only negligible differences in migrated images compared to full-precision implementations, while achieving substantial gains in memory and disk efficiency. Further analysis explores the feasibility of using FP16 for computation, revealing that while performance improves, the numerical errors are too large for production-quality imaging. These results suggest that FP16 is highly effective for storage and communication in RTM workflows, but not yet suitable for core numerical operations in production-scale deployments.