A 22nm Resource-Frugal Hyper-Heterogeneous Multi-Modal System-on-Chip Towards In-Orbit Computing

Integrating artificial intelligence (AI) into in-orbit computing offers significant benefits, but current satellites face challenges in processing large sensor data volumes due to limited communication and computing resources, resulting in high latency [1]. Intelligent Early Discard (IED) [2] addresses this by filtering irrelevant data early, optimizing bandwidth and data usage. However, this demands highperformance onboard computing for efficient data preprocessing and AI acceleration [3], [4]. Additionally, Space radiation, including solar energetic particles and cosmic rays, can cause Single Event Upsets (SEUs) in satellite systems [5], risking mission failure and increasing reliability demands [6], [7]. To tackle these challenges, we propose a resource-frugal hyper-heterogeneous System-on-Chip (SoC) architecture for in-orbit computing. The SoC features two modes: (1) a specialized computation engine for AI acceleration, and (2) a multicore mode with dual-core lock-step (DCLS) and vector computing for efficient, fault-tolerant data processing (Fig. 1). This resource-frugal architecture enables full sharing of Processing Elements (PEs) and memories for dynamic workload allocation, enhancing in-orbit performance by processing IED data directly on the satellite and reducing costly data transmission to Earth.