Performance analysis of orbit propagation on GPU architecture

The development of space technology and the diversification of the scope of space missions are leading to an increasing complexity of on-board and off-board computational operations. Applications such as real-time orbit determination or astrodynamics simulations for mission analysis require high-performance and efficient software and hardware interfaces. Already widely used in computer vision, big data analysis, and scientific simulations, Graphic Processing Unit (GPU) technology has been considered as a possible architecture to address this problem. For this reason, this work aims to understand the extent to which parallel computation on GPU architectures can improve orbit propagation performance compared to traditional CPU-based methods. To assess the performance of the two approaches, a Monte Carlo analysis of trajectory propagation and its uncertainty has been conducted using custom software built in both a serial-flow version and a parallel-flow version running respectively in C++ and C++/CUDA. The propagation outcomes are compared to those from the same simulations run in ESA-Godot, Orekit, and GMAT mission analysis tools, by monitoring the corresponding computational times. A sample of the satellite Sentinel-3A's trajectory is taken as a reference test case scenario. The hardware used includes a high-performance desktop device, a consumer desktop device, and a low-power embedded system. Results show that GPU parallelization can achieve speedups of up to two orders of magnitude for large-scale propagation, while CPU-based architectures remain more efficient for the propagation of a small number of trajectories. Therefore, GPU architectures provide a valuable alternative for addressing the heavy computational demands of orbit propagation in modern space mission applications.