E-NERVE: An Edge-Adaptive, NAF-Empowered, Resource-Efficient Versatile AI Inference Engine

Abstract

Non-linear activation functions (NAFs) account for only a small portion of deep neural network compute, yet provisioning dedicated activation hardware can increase area and create dark silicon in edge accelerators. This letter presents E-NERVE, a two-stage, runtime-adaptive inference engine that decouples MAC/FMA computation from activation evaluation using a shared MAC bank and a single time-multiplexed CORDIC-based activation core (V-ACT). V-ACT supports ReLU, Sigmoid, Tanh, and Softmax through runtime selection, while a PISO interface overlaps serial activation with the next tile’s MAC execution to avoid activation-induced throughput loss. Results show that activation reuse and MAC sharing reduce resource and energy overhead with accuracy degradation under 2.7% for selected models evaluated on CIFAR-100/ImageNet. For a 196:64:32:32:10 network (MNIST), the VC707 prototype consumes 0.267 W, while for a 28 nm ASIC, E-NERVE consumes 1.26 mm2, 0.38 W and achieves 6.78 GOPS/W with 98.32% accuracy. A Pynq-Z2 prototype achieves 154.6 ms latency at 0.25 W, validating the proposed latency-hiding activation reuse for resource-constrained edge inference. © 2009-2012 IEEE.