Energy Adaptive Reconfigurable Neural Accelerator Architecture for Efficient Edge Artificial Intelligence Applications

The growing deployment of artificial intelligence (AI) applications on edge devices presents critical challenges in balancing energy efficiency, thermal stability, and computational performance. Conventional hardware accelerators are not always dynamically adjustable to changing workloads and environmental conditions, leading to power waste and low model performance. In the given work, we introduce a new AI-based Energy Adaptive Reconfigurable Neural Accelerator Architecture, which is based on multi-dimensional entropy profiling, parametric synchronous dataflow modeling, and approximate bitwidth folding to smartly manage both resource usage and energy consumption. EARN-AI is able to provide high throughput and low latency inference with dynamic tile activation, thermal-aware scheduling, and hybrid non-volatile memory integration even operating with a limited amount of power and temperature. Experimental computation with a wide variety of workloads including MobileNetV2, ResNet18, and YOLOv4-Tiny on edge computing systems (Jetson Nano, Raspberry Pi 4, FPGA) show that energy consumption can be reduced by 35-50 %, and no significant loss of model accuracy (worse than 2.5% degradation at 6-bit precision) occurs. EARN-AI is more efficient in power consumption and flexibility as compared to state-of-the-art accelerators such as Eyeriss and EdgeTPU. The architecture is a real-time, battery-sensitive AI application, framework-compatible, and scalable architecture. These findings are important to highlight that it could be a universal edge AI inference engine of the next generation smart devices.