Tightly-Coupled FPGA Accelerator for Molecular Dynamics Simulation: Hardware-Software Co-Design and Fine-Grained Task Management
Zekang Cheng, Zerong He, Xi Jin · 2024
Over the past decades, Molecular Dynamics (MD) has been extensively utilized for drug design, protein structure prediction, and system analysis in computational chemistry and biology. However, previous acceleration efforts have often faced challenges with either host-device execution modes, leading to costly communication overheads, or complete FPGA implementations that sacrifice flexibility and programmability. In this paper, we present a novel tightly-coupled MD execution framework, combining a single FPGA with a Hard-core CPU, resulting in an impressive 10x performance improvement compared to state-of-the-art CPU implementations (e.g., Gromacs). Our system exhibits characteristics reminiscent of a fine-grained CPU-based MD execution with Low latency computation accelerators. To achieve such efficiency, our hardware-software co-design approach emphasizes reducing scheduling expenses through a decentralized dependency management strategy and a hardware-assisted Multi-Producer Multi-Consumer (MPMC) queue. Importantly, our methodology is not limited to MD applications alone but can be readily applied to a wide range of fine-grained task-based applications. Moreover, the MPMC queue has the potential to scale and be adapted for use in FPGA clusters of larger scale, further extending its applicability and relevance in diverse computing environments.