Scalable Hydrodynamics on multiple Field-Programmable Gate Arrays (FPGAs)

This paper presents scalable 2D and 3D Hydrodynamics solver implementations on FPGAs using Intel’s oneAPI framework, addressing challenges in porting stencil-based computations from CPU/GPU architectures on FPGAs. FPGAs offer customizable hardware with on-chip memory that can be custom-tailored for High Performance Computing (HPC) applications. Despite a much lower throughput, FPGAs achieve comparable energy efficiency with domain decomposition compared to NVIDIA A100 GPU. Optimizations leverage FPGA’s large cache to minimize data streaming and ensure data reuse, reducing memory bandwidth. Domain decomposition enabled pipeline duplication for scalable single FPGA workload, but also the distribution of the work on two Agilex 7 cards. This work demonstrates FPGA viability for computationally intensive simulations, contributing to scalable algorithms for heterogeneous HPC systems and offering insights for exascale computing.