Acceleration of deep learning applications for astrophysics on FPGAs

Abstract

Deep learning algorithms have been widely used in the past decade due to their effectiveness and robustness in information processing in various scientific domains. With the evolvement of large-scale datasets and deep learning applications used for gravitational wave astrophysics and large-scale electromagnetic surveys, an accessible and efficient framework is essential to accelerate domain-specific applications on acceleration platforms. Although CPUs and GPUs are common platforms for training and inferencing deep learning applications given their ease-of-use and support for popular deep learning frameworks, field-programmable gate array (FPGA)-based accelerators have rapidly gained their popularity because of their capacity to deliver superior performance for real-time applications with relatively low power consumption. In this thesis, we investigate the process of porting deep learning algorithms developed for Multi-Messenger Astrophysics and particle physics on FPGA accelerators and propose a cyberinfrastructure that enables the use of FPGAs to accelerate algorithms inference. Also, we explore and evaluate three main development stacks, hls4ml, Vitis AI, and TVM, used for deploying deep neural network (DNN) models on cloud FPGA-based hardware accelerators.