Hardware-Aware NAS with Post-Training Quantization for Edge AI

A new arXiv paper (2608.13293) proposes a three-stage pipeline for hardware-aware Neural Architecture Search (NAS) that integrates post-training quantization (PTQ) and automated hardware exploration for edge AI deployment. The first stage uses a hardware-agnostic Pareto rank surrogate on NAS-Bench-201 to efficiently rank architectures. The second stage applies a quantization bridge with Pareto-aware filtering and feedback control, addressing the lack of analytical attention on PTQ effects on NAS-discovered Pareto structures. The third stage employs an evolutionary domain-specific accelerator exploration to map quantized architectures onto reconfigurable hardware. The paper aims to decouple architecture and quantization design, reducing search complexity while maintaining accuracy, efficiency, and hardware deployability. It highlights the gap in existing methods that tightly couple architecture and quantization, and proposes a simpler post-search quantization strategy. The work is relevant to the growing field of edge AI, where neural networks must be simultaneously accurate, computationally efficient, and hardware-deployable.