Energy-Efficient Approximate Edge Inference Systems

The rapid proliferation of the Internet Things and dramatic resurgence artificial intelligence based application workloads have led to immense interest in performing inference on energy-constrained edge devices. Approximate computing (a design paradigm that trades off a small degradation quality for disproportionate energy savings) is promising technique enable energy-efficient at edge. This article introduces concept an approximate system ( AxIS ) proposes systematic methodology perform joint approximations between different subsystems deep neural network (DNN)-based system, leading significant benefits compared approximating individual isolation. We use smart camera executes various DNN-based image classification object detection applications illustrate how sensor, memory, compute, communication can all be approximated synergistically. demonstrate our proposed using two variants system: (a) Cam Edge , where DNN executed locally device, (b) Cloud device sends captured remote cloud server DNN. prototyped such Intel Stratix IV GX-based Terasic TR4-230 FPGA development board. Experimental results obtained six large DNNs four compact running savings (≈ 1.6× -4.7× ≈ 1.5× -3.6× DNNs), minimal (<1%) loss application-level quality. Furthermore, exhibit -5.2× similar loss. Compared single subsystem isolation, achieves 1.05× -3.25× gains 1.35× -4.2× average,