Toward Reliable Artificial Intelligence Acceleration in Space: Design and Optimization of Onboard Processing Architectures

Artificial Intelligence is being increasingly considered for a wide range of applications in space, including Earth Observation tasks and future autonomous missions. Executing Machine Learning tasks directly onboard offers significant advantages in bandwidth, latency, and operational costs. Despite this, the adoption of AI, especially in critical scenarios, is hindered by limited onboard resources, power budget, and stringent reliability requirements. ML algorithms are computationally intensive, requiring dedicated hardware accelerators that must provide high performance while meeting space system constraints. Reliability threats, such as Single Event Effects, pose major challenges and can compromise mission success, thus requiring robust strategies at both hardware and system levels. This work proposes strategies for dependable AI execution across low and high-criticality scenarios. As a short-term solution, the soft GPGPU paradigm emerges as a promising solution. It combines the flexibility and flight heritage of FPGAs with the computational strengths of GPUs to meet diverse mission requirements. As such, this thesis focuses on a soft GPU IP core architecture, namely GPU@SAT from IngeniArs S.r.l., detailing the reliability-enhancing strategies implemented and the development of a fault-tolerant SoC based on the IP. Looking beyond short-term solutions, the thesis also fosters the development of next-generation, reliable, and energy-efficient ASICs for space systems. To this end, the CGR-AI Engine is introduced as a novel CGRA-based processing platform that combines the programmability of a RISC-V with the reconfigurability of a CGRA processing matrix.