An FPGA-based efficient accelerator for fault interaction of rupture dynamics
摘要
Efficiently predicting aftershocks based on rupture dynamics simulation is a crucial task in high-performance computing, traditionally dependent on supercomputers. However, the constraints of power supply makes supercomputers impractical right after a primary earthquake event. This positions FPGAs, known for their high power efficiency and reconfigurability, as a highly promising alternative. Within rupture dynamics simulation, fault interaction is the most computationally intensive component, making its acceleration crucial for enhancing overall performance. By thoroughly considering both the computational characteristics of fault interaction and the reconfigurable capabilities of FPGAs, we have devised a high-performance, power-efficient accelerator for fault interaction. On one hand, we conduct an in-depth analysis of the algorithm’s data dependencies and exploit parallelization at multiple levels to maximize performance. On the other hand, we propose novel dataflow optimizations, such as prefetching and overlapping computations across different stages, to further enhance efficiency. Additionally, we implement a latency-matching strategy and a flag-based mechanism to ensure seamless coordination between computational stages. Experimental results demonstrate the superior efficiency of our FPGA accelerator across geological models. Against a 12-core Intel Xeon CPU, the FPGA achieves 12.3