Paper Key : IRJ************348
Author: Muhammad Nouman
Date Published: 21 Nov 2024
Abstract
ABSTRACT In the current era of rapid growth in Internet of Things (IoT) devices and the demand for low power, high efficiency computation solutions in edge computing have has encouraged research efforts in alternative arithmetic suitable for resource-constrained environments. This paper investigates the implementation in fixed-point arithmetic over edge devices to achieve more computational efficiency and comparatively less power consumption compared to floating-point arithmetic. Experiments were performed on three different hardware platforms which are PIC32MZ2048EFG, Node MCU 8266 and Raspberry Pi zero. The findings show that fixed-point arithmetic consumes 50% less energy and executes 66% faster in terms of processing speed with minor accuracy trade-offs. The paper further demonstrates that two or threefold reductions in training time and memory usage can be achieved at the expense of acceptable accuracy loss when examining the performance of MNIST dataset neural network models in fixed-point versus floating point computation. These findings of this paper highlight a significant lower cost on edge computing is reached via fixed-point instead of floating-point that allows AI deployment on devices with limited computation and storage resources.Keywords: edge computing, floating-point arithmetic, fixed-point arithmetic, neural network models, AI.