A Near-Threshold Machine Learning Accelerator for Energy-Efficient Edge Inference

NTC-EdgeMAC is a near-threshold machine learning accelerator designed for high energy efficiency in edge inference. The architecture integrates several coordinated techniques, including a heterogeneous mixed-precision MAC array optimized for near-threshold operation, adaptive bodybias control with dynamic voltage and frequency scaling to address process and temperature variations, and in-situ timingerror detection with lightweight correction for improved resilience. Workload-aware optimizations such as early-exit processing and per-layer precision adjustment further reduce energy consumption. A cycle-accurate RTL model synthesized in a 28-nm CMOS process demonstrates significant reductions in energy and energy‒delay product for models such as MobileNetV2, with accuracy loss maintained below 0.5%. Detailed descriptions of the hardware architecture, control mechanisms, and comparative performance with recent lowpower accelerators are presented.