A Methodology for Evaluating and Analyzing FPGA-Accelerated, Deep-Learning Applications for Onboard Space Processing

Due to continued innovations in onboard data analysis and spacecraft autonomy, enabled by deep learning (DL), modern spacecraft require dependable, high-performance computers to process onboard an immense volume of raw sensor data into actionable information to formulate critical decisions autonomously. To enable compute-intensive DL algorithms, commercial-off-the-shelf processors, including FPGAs and system-on-chips, are often employed for their superior performance, energy-efficiency, and affordability compared to traditional radiation-hardened alternatives; however, these processors are highly susceptible to radiation-induced single-event effects (SEEs) that can degrade the dependability of DL applications. Researchers have created a diverse collection of DL models that perform a variety of tasks useful for Earth-observation missions. However, due to characteristic differences between models and accelerators, their tradeoffs can vary in terms of accuracy, area, performance, energy-efficiency, and dependability, which are factors crucial for resource-constrained and mission-critical systems. To select the optimal DL solution that maximizes inference performance, conserves onboard resources, and satisfies mission dependability requirements, a methodology is required to evaluate and compare the tradeoffs between competing options. In this paper, we propose a methodology for evaluating and analyzing the tradeoffs of FPGA-accelerated DL models, including a hierarchical fault-injection approach to accelerate the characterization of SEE susceptibility of DL solutions in terms of well-established dependability metrics. Furthermore, we identify performance and dependability trends, analyze the impact of SEEs on the inference accuracy, and predict design fault rates for near-Earth orbital environments. To demonstrate the versatility of our methodology, we evaluate and analyze four semantic-segmentation models accelerated on four Xilinx Deep-Learning Processing Unit accelerators.