A Cross-Layer Survey of Low-Power Machine Learning From Algorithm to Hardware

Abstract

Energy is now a primary constraint for modern machine-learning training and inference, from hyperscale data centers to battery-constrained edge devices. Prior surveys cover efficient and green deep learning, edge artificial intelligence (AI), accelerator architectures, lifecycle-level energy reduction, and measurement or benchmarking methodology, but often treat these threads separately. This survey complements them with a physically grounded, measurement-aware energy-accounting lens connecting Model Architecture & Computation, Precision & Numerics, Input & Dataset, System & Compiler (Runtime), and Hardware & Physics Substrate decisions to operation count, per-operation cost, data movement, and platform/static overheads. A recurring pattern is that robust energy savings arise when arithmetic-work reductions are coupled with lower data movement and explicit measurement boundaries; proxy-only gains can disappear when sparse metadata, conversion, communication, or platform overheads are included. We emphasize task-normalized reporting, lifecycle-aware amortization of optimization overheads, and claim-matched measurement methods. Quantitative results are drawn under a documented, scope-driven selection and normalized in evidence tables recording platform, measurement boundary, and whether each figure was metered, modeled, or simulated; a dedicated threats-to-validity section states what the resulting synthesis does and does not support. Coverage spans tiny machine learning (TinyML) and on-device training through federated and edge–cloud execution, and concludes with open challenges in unified toolchains, reporting rigor, and emerging accelerator directions.