A High-Efficiency and Low-Storage Algorithm for Seismic Simulation Using Half Precision and Scalable Vector Extension on ARM Platforms

Seismic simulations are essential for ground motion characterization and seismic hazard mitigation. However, achieving accurate seismic modelling requires highly refined computational grids, which impose severe memory and computational challenges. Traditional seismic solvers based on single-precision floating-point 32-bit (FP32) arithmetic, suffer from excessive memory consumption, low-memory access efficiency and limited computational efficiency. In contrast, half-precision floating-point 16-bit (FP16) halves memory usage and effectively doubles memory access efficiency, making it attractive for large-scale seismic simulations. However, direct application of FP16 to classical elastic wave equations is challenging due to overflow and underflow caused by the wide dynamic range of physical variables. In this work, we reformulate the elastic wave equations by introducing three dimensionless scaling constants, Cv, Cs, and Cp, and derive an FP16-based elastic wave equation. Furthermore, we provided a practical strategy for determining these constants based on the source time function, ensuring that velocity and stress variables remain within the representable range of FP16. To maintain FP32-level accuracy, a mixed-precision strategy using “FP16 storage and FP32 arithmetic” is adopted. From a computational perspective, we further exploit the Scalable Vector Extension (SVE) on ARM architectures to accelerate stencil-based computations. However, effectively combining FP16 with SVE introduces additional challenges, including stencil restructuring for vectorization and data layout mismatches arising from “FP16 storage and FP32 arithmetic”. To overcome these challenges, this study develops three complementary seismic solvers on the ARM architecture: an FP16-based solver, an SVE-accelerated solver, and an FP16–SVE hybrid solver that integrates memory efficiency with vectorized computation. All three solvers are implemented, systematically validated, and benchmarked using both synthetic test cases and real earthquake simulations. Numerical results demonstrate near-identical agreement with a reference FP32 solver across diverse seismic scenarios. In particular, the FP16–SVE hybrid solver reduces memory consumption by approximately 50% and achieves up to a threefold speedup, delivering more than a 2.3× acceleration in real-world earthquake simulations. These results highlight the strong potential of the proposed FP16–SVE approach for enabling large-scale, high-efficiency, and near-real-time seismic simulations and earthquake hazard assessment on ARM-based platforms.