Abstract Modern edge devices, such as cameras, drones, and internet-of-things nodes, rely on machine learning to enable a wide range of intelligent applications. However, deploying machine learning models directly on the often resource-constrained edge devices demands substantial memory footprints and computational power for real-time inference using traditional digital computing architectures. In this paper, we present WISE, computing architecture for wireless edge networks with two key innovations: disaggregated model access via over-the-air wireless broadcasting for simultaneous inference on multiple edge devices, and in-physics computation of general complex-valued matrix-vector multiplications directly at radio frequency driven by a single frequency mixer. Using a software-defined rad…
Abstract Modern edge devices, such as cameras, drones, and internet-of-things nodes, rely on machine learning to enable a wide range of intelligent applications. However, deploying machine learning models directly on the often resource-constrained edge devices demands substantial memory footprints and computational power for real-time inference using traditional digital computing architectures. In this paper, we present WISE, computing architecture for wireless edge networks with two key innovations: disaggregated model access via over-the-air wireless broadcasting for simultaneous inference on multiple edge devices, and in-physics computation of general complex-valued matrix-vector multiplications directly at radio frequency driven by a single frequency mixer. Using a software-defined radio platform, WISE achieves 95.7% image classification accuracy (97.2% audio classification accuracy) with ultralow energy consumption of 6.0 fJ/MAC (2.8 fJ/MAC), which is more than 10× improvement compared to traditional digital computing, e.g., on modern GPUs.