From NAS to Hardware: Deploying Quantized Automatic Modulation Classification Models on FPGA for 6G Edge Intelligence
Future 6G edge-intelligent radios require neural networks that satisfy strict latency and energy constraints, yet the practical behaviour of quantization-aware Neural Architecture Search (NAS) models on real Field-Programmable Gate Array (FPGA) hardware remains insufficiently explored. This work evaluates a set of Pareto-optimal architectures generated by a quantization-aware NAS MONAS-LQ, together with state-of-the-art reference models, when deployed on an FPGA using the Brevitas-FINN-Vivado toolchain. The results show that all models found by MONAS-LQ maintain accuracy within 0.30% of server-side execution, with several exhibiting slight improvements, while energy consumption remains below 20 mJ/sample and latency is dominated by early convolutional layers rather than overall model size. Compared to existing quantized architectures, the MONAS-LQ models achieve more favourable accuracy–efficiency trade-offs. These findings highlight the relevance of hardware-aware NAS for deriving deployable and energy-efficient Deep Learning (DL) models tailored to the resource constraints of future edge-intelligent radio systems.