Acceleration of ANN based molecular dynamics using GPU and FPGA

Abstract

High Performance Computing (HPC) technology uses large number of powerful processors, working in parallel, to process large amounts of data and solve complex problems at very high speeds. Processors used in the HPC systems were primarily Central Processing Units (CPUs). As the data became multidimensional, the use Graphics Processing Units (GPUs) as co-processors for CPUs in HPC systems became popular making the HPC systems heterogeneous. GPUs process data parallelly with higher throughput than the CPUs due to the large number of processing cores. CPUs have higher per-core throughput compared to the GPUs. Keeping this in mind, the workload in a heterogenous computing system is divided between the different processors in such a way to provide maximum throughput. Field Programmable Gate Arrays (FPGAs) provide an architecture with large number of programmable hardware components which facilitates implementing various logic on the hardware. This project deals with accelerating ANN based Molecular Dynamics calculations by utilizing the hardware accelerators, GPU and FPGA. In the first phase of the project, a heterogeneous computing system with one CPU and one GPU is used to calculate the energy and force on each atom from their positions in a Au nanocluster. The computed forces values are used to calculate the new coordinates of atoms, thereby forming a closed loop. CUDA Programming Model is employed to program the workload onto the Heterogeneous Computing System. The sequential computations of the workload are processed on the CPU and the parallel computations of the workload are offloaded to the GPU for processing. This approach of offloading along with effectively utilizing the GPU cores provides better throughput than computations on CPU system.