Neural Architecture Search for Energy-Efficient Inference on Heterogeneous Edge Platforms

Deploying deep neural networks on resource-constrained edge devices requires careful trade-offs between accuracy, latency, memory footprint, and energy consumption—objectives that vary significantly across heterogeneous hardware platforms. We present HeteroNAS, a neural architecture search framework that explicitly targets multi-objective optimization across CPUs, mobile GPUs, NPUs, and FPGAs simultaneously. HeteroNAS employs a hardware-aware supernet with platform-specific operation libraries and a multi-fidelity Bayesian optimization algorithm to efficiently explore the architecture space. Evaluated on ImageNet classification and VOC2012 object detection, HeteroNAS discovers architectures achieving Pareto-optimal accuracy-efficiency trade-offs, reducing inference energy by 2.8–4.6× compared to MobileNetV3 at equivalent accuracy.