Implementation of quasi-Newton algorithm on FPGA for IoT endpoint devices

With the recent developments in the internet of things (IoT), there has been a significant rapid generation of data. Theoretically, machine learning can help edge devices by providing a better analysis and processing of data near the data source. However, solving the nonlinear optimisation problem is time-consuming for IoT edge devices. A standard method for solving the nonlinear optimisation problems in machine learning models is the Broyden-Fletcher-Goldfarb-Shanno (BFGS-QN) method. Since the field-programmable gate arrays (FPGAs) are customisable, reconfigurable, highly parallel and cost-effective, the present study envisaged the implementation of the BFGS-QN algorithm on an FPGA platform. The use of half-precision floating-point numbers and single-precision floating-point numbers to save the FPGA resources were adopted to implement the BFGS-QN algorithm on an FPGA platform. The results indicate that compared to the single-precision floating-point numbers, the implementation of the mixed-precision BFGS-QN algorithm reduced 27.1% look-up tables, 18.2% flip-flops and 17.9% distributed random memory. Copyright © 2022 Inderscience Enterprises Ltd.