RISC-V-Based heterogeneous computing platform for satellite and space applications

Space systems are rapidly evolving into complex Cyber-Physical Systems, where computation, physical processes, and control systems closely interact. These systems, including satellites and launch vehicles, demand high real-time processing capabilities, reliability, and resilience to withstand the harsh space environment, where radiation risks significantly challenge electronic components. Balancing computational efficiency with fault tolerance is crucial in Space Cyber-Physical Systems, placing rigorous demands on their computing architectures. Modern data-driven workloads, including machine learning and artificial intelligence, are becoming increasingly central to space missions, imposing stringent performance and energy-efficiency requirements on onboard computers. Traditional CPUs play an essential role in control and decision-making tasks due to their programmability but lack the capacity to handle the intensive computations required by contemporary AI algorithms. This performance gap, especially in space applications, has led to exploring alternative, more efficient computing paradigms to meet these demands. Heterogeneous systems, integrating general-purpose CPUs with specialized accelerators, present a promising approach to address this challenge. The open-source RISC-V Instruction Set Architecture enables the development of custom, adaptable, and energy-efficient architectures, ideal for mission-specific requirements and fault mitigation strategies in Space Cyber-Physical Systems. While heterogeneous systems are prevalent in consumer markets, their potential in space applications remains underexplored. This thesis investigates a RISC-V-based heterogeneous platform designed for space, introducing key contributions: a multicore cluster with a Hybrid Modular Redundancy scheme, the RedMulE accelerator for efficient neural network training, and Astral, a System-on-Chip featuring a RISC-V-based control domain, specialized accelerators, extensive peripherals, and a hardware root of trust, all designed for secure, energy-efficient execution of AI workloads in space applications.