Satellite on-board processing is experiencing significant growth across diverse sectors, including telecom, Earth observation, navigation, exploration, and science. While current on-board processing techniques mainly rely on conventional methods such as filtering, FFT, encoding, and modulation, there is a rising interest in harnessing machine learning and neural network approaches to tackle complex challenges such as pattern recognition and image segmentation. Moreover, the expansion of edge processing, which takes place at the data generation point—in this case, satellites—is predicted to see rapid development in the coming years.
The HPDP-40 is an exceptional radiation-hardened and high-performance co-processor specifically designed for space applications. This solution distinguishes itself with its remarkable features, including low power consumption, exceptional flexibility, on-the-fly re-programmability, compact form, and low mass. It opens up new possibilities for advanced data processing in space.
Successful testing has been conducted on Artificial Intelligence (AI) using the HPDP, albeit with a limited scope due to the lack of a generic AI framework tailored for this co-processor. However, a groundbreaking solution is on the horizon.
Klepsydra software products present a transformative opportunity by deploying them on widely used processors and operating systems, resulting in a substantial tenfold increase in data processing rate and an impressive 50% reduction in energy consumption. At the core of Klepsydra’s offerings is the Streaming Distribution Optimizer (SDO), an integral part of Klepsydra AI. The SDO employs Genetic Algorithms (GA) to identify the optimal distribution of computing resources for executing AI models on the target processor, taking into account critical factors such as latency, power consumption, and data rate.
A pioneering project, code name MANDALA, is now underway to adapt Klepsydra AI and SDO specifically for AI execution on the HPDP, addressing the existing scarcity of space-qualified AI inference solutions. This ambitious endeavor, scheduled to commence on June 1, 2023, and expected to conclude on May 31, 2024, aims to overcome the technical challenges arising from the unique architecture of the HPDP co-processor and its programming language, NML. The integration of AI applications into this cutting-edge solution is a priority focus area, albeit not without risks.
The primary advantage of this groundbreaking solution lies in its significant simplification of AI model deployment on the HPDP. Klepsydra ensures compatibility with major AI frameworks, eliminating the need for HPDP-specific coding. As a result, this innovation holds the potential to enable AI applications in future missions by the European Space Agency (ESA) that demand high availability, including landers and deep space robotics. Ultimately, this project could establish a fully European hardware/software stack, reinforcing Europe’s capability to leverage AI in space.