System-Level Modeling of GPU/FPGA Clusters for Molecular Dynamics Simulations

FPGA-accelerated molecular dynamics (MD) research dates back to almost two decades ago and is still being actively studied. MD on FPGA clusters, however, is still in its initial phase with only small systems built and limited performance studies. Given the cost of building accelerator clusters, and (as we show) the number of plausible architectures, a thorough study is needed. In particular, we investigate both FPGA and GPU/FPGA hybrid clusters. The latter are potentially attractive given the broad availability of GPU clusters and use of GPUs for MD, but the current inability of GPUs to scale for certain critical domains. In this work, we model four promising MD accelerator platforms, including FPGA-only systems with homogeneous and heterogeneous nodes, an existing FPGA-GPU hybrid system (the Cygnus supercomputer), and a synthesis of the commercially available Nvidia DGX-1/DGX-2 products with an FPGA cluster. The models are compared and evaluated, and we find that each of the platforms is preferable for some domains.