Edge-Native Graph Transformer Networks for Real-Time RIS Beamforming on Low-Cost Hardware

Abstract

Real-time Reconfigurable Intelligent Surface (RIS) beamforming requires low-latency inference that is difficult to satisfy using cloud-based or GPU-based infrastructures at the network edge. Existing studies lack simultaneous deployment and benchmarking of competing neural-network architectures as autonomous, closed-loop RIS controllers on low-cost single-board computers (SBCs). This paper deploys four architectures (a physics-aware Graph Transformer Network (GTN), a Residual MLP, a Deep Reinforcement Learning actor, and an Algorithmic Unfolding network) natively on a Raspberry Pi 4B. Models are evaluated under a 50/50 multi-objective framework that equally weights RF fidelity and computational feasibility. The proposed GTN achieves a 29.49 dB mean SNR (within 0.50 dB of the Simulated Annealing oracle) with 39.7 ms inference latency and 5.37 MB memory footprint, demonstrating the viability of SBCs as viable low-cost platforms for autonomous RIS control. The Accuracy Paradox reveals that purely supervised Binary Cross-Entropy (BCE) is insufficient for 1-bit phase optimization.