The rapid expansion of satellite constellations has led to an unprecedented surge in data generation, highlighting the escalating demand for high-performance computing over space. A recent article titled “Computing over Space: Status, Challenges, and Opportunities” published in Engineering delves into the current state, challenges, and future prospects of space computing.

The article, authored by Yaoqi Liu, Yinhe Han, Hongxin Li, Shuhao Gu, Jibing Qiu, and Ting Li, outlines how the increasing resolution of remote-sensing images and the limited bandwidth of satellite–ground communications necessitate onboard computing to save transmission and processing time. For instance, as the ground resolution of remote-sensing images has improved from 10.0 to 0.3 m, the data volume has increased by approximately 1000 times. Onboard computing can extract high-value information from this vast amount of data, significantly reducing the required transmission bandwidth and service time.

Beyond remote sensing, satellite communication is another critical application of space computing. The onboard deployment of core networks is essential for realizing satellite networks, with space computing performance being a key limiting factor for tasks such as signal processing, multiplexing, traffic management, and resource allocation. Prof. Shangguang Wang from Beijing University of Posts and Telecommunications has proposed the Tiansuan Constellation and conducted experiments on the onboard deployment of the core network.

The article also discusses the challenges of using commercial off-the-shelf (COTS) devices in space computing. While COTS devices have improved the performance of space computing, there is still a significant gap between current space computing systems and the most advanced ground-based systems. For example, the RAD5500 processors commonly used in space have a performance of only 0.9 giga floating-point operations per second (GFlops), compared to the NVIDIA A100, which achieves 156 tera floating-point operations per second (TFlops).

To address these challenges, the authors propose a multi-level fault-tolerant system, including component-level, system architecture-level, software-level, and algorithm-level solutions. At the component level, methods such as instruction-level time redundancy and multi-device redundancy can improve error-correction capability. At the system architecture level, key module redundancy, cold/hot backup, and watchdog mechanisms can increase tolerance to critical component failures. At the software level, technologies like cloud-native or microkernel can enhance the availability of the operating system. Finally, at the algorithm level, redundancy can be applied to data or neural network models to reduce silent data corruption.

Thermal control is another critical aspect of space computing. In the vacuum of space, with its extreme temperature differences, effective heat dissipation is crucial. The article proposes a hybrid passive–active cooling (HPAC) method, where the active cooling part is responsible for cooling high-power chips, while the passive cooling part handles low-power chips. This ensures basic functionality even if the fluid loop fails.
On the application front, the integration of large language models (LLMs) into satellite systems offers a promising solution for achieving intelligent information fusion and natural language interpretation of human instructions. The authors have implemented a visual large language model (VLLM) on the Jiguang 1000-OSE platform, demonstrating the potential for bidirectional natural language communication with ground operators and automated analysis of remote-sensing images.

High-performance computing over space is poised to revolutionize fields such as autonomous exploration and space-based data processing. By addressing key challenges in computing architecture, thermal control, and applications, space computing can enable real-time data processing and analysis across diverse fields.

The paper “Computing over Space: Status, Challenges, and Opportunities,” is authored by Yaoqi Liu, Yinhe Han, Hongxin Li, Shuhao Gu, Jibing Qiu, Ting Li. Full text of the open access paper: https://doi.org/10.1016/j.eng.2025.06.005. For more information about Engineering, visit the website at https://www.sciencedirect.com/journal/engineering.