Distributed Deep Learning Inference on DPU-Based FPGA Systems for On-Board Earth Observation

The growing adoption of Artificial Intelligence is driving demand for accurate, robust models, often characterized by deep, large-scale neural networks. While these models achieve high performance, their computational cost leads to significant inference latency, particularly on embedded, resource-constrained platforms such as FPGA-based architectures. This challenge is especially relevant in the space domain, where on-board deep learning is required for data analysis and decision-making under strict constraints on latency and computational resources. To address this challenge, this work proposes a distributed inference architecture based on FPGA accelerators, combining Deep Processing Units with a cluster-oriented execution model over TCP. The approach leverages a high-level development flow based on Vitis AI, enabling deployment without low-level hardware design expertise. Inference is executed across multiple FPGA boards, overcoming the limitations of a single device while maintaining portability and ease of integration. The proposed solution is validated on a cloud-detection task using satellite imagery, representative of realistic on-board scenarios. Experimental results show that the proposed approach achieves a speedup of up to 1.93× over single-board execution, demonstrating its effectiveness for large neural networks and suitability for on-board processing in space applications.