A Hardware Pipeline with High Energy and Resource Efficiency for FMM Acceleration

The fast multipole method (FMM) is a promising mathematical technique that accelerates the calculation of long-ranged forces in the large-sized n-body problem. Existing implementations of the FMM on general-purpose processors are energy and resource inefficient. To mitigate these issues, we propose a hardware pipeline that accelerates three key FMM steps. The pipeline improves energy efficiency by exploiting fine-granularity parallelism of the FMM. We reuse the pipeline for different FMM steps to reduce resource usage by 66%. Compared to the state-of-the-art implementations on CPUs and GPUs, our implementation requires 15% less energy and delivers 2.61 times more floating-point operations.