This work presents the initial assessment of the performance and power needs of two commercially available Systems-on-a-Chip (SoC) featuring both Central Processing Units (CPU) and Graphical Processing Units (GPU): the NVIDIA Jetson Nano Developer Kit and the NVIDIA Jetson AGX Orin Developer Kit. The objective of this evaluation is to offer an early estimation of the GPUs’ performance and technical viability for deploying on-board machine learning tasks in an on-board processing subsystem. This evaluation establishes a baseline for future optimization of on-board processing within resource-limited environments, specifically nanosatellite systems. Monitoring processes are configured to obtain a continuous observation of parameters such as execution and training time, CPU and GPU usages, throughput, performance, components power consumption, temperatures and efficiency. Benchmarking different conditions can provide results useful to determine the requirements, criteria and trade-offs to be considered when implementing such devices. Preliminary findings confirm the feasibility of using the NVIDIA Jetson family of devices for space applications involving demanding data processing or Artificial Intelligence (AI) and Machine Learning (ML) models. Additional emphasis has been put in tailoring the Operating System (OS); by eliminating unnecessary processes irrelevant to the desired pipeline or the tuning of resources through the different power modes is possible to alleviate the requirements on the power and thermal subsystems. Finally, an interest in training those AI/ML models in space is taken: the study case of this work, a Water Electrolysis Propulsion (WEP) controller, requires the ability to retrain the neural network in order to autonomously operate the spacecraft. Considering different configurations of the SoCs’ resources has lead to the conclusion that training is also viable.
«
Login
BibTeX