An OpenCL 3D FFT for Molecular Dynamics Distributed Across Multiple FPGAs

3D FFTs are used to accelerate MD electrostatic forces computations but are difficult to parallelize due to communications requirements. We present a distributed OpenCL 3D FFT implementation on Intel Stratix 10 FPGAs for grids up to {\boldmath $128^3$}. We use FPGA hardware features such as HBM2 memory and multiple 100 Gbps links to provide scalable memory accesses and communications. Our implementation outperforms GPUs for smaller FFTs, even without distribution. For {\boldmath$32^3$} we achieve 4.4 microseconds on a single FPGA, similar to Anton 1 on 512 nodes. For 8 parallel pipelines (hardware limited), we reach the same performance both locally and distributed, showing that communications are not limiting the performance. Our FFT implementation is designed to be part of the electrostatic force pipeline of a scalable MD engine.