Detecting Handwritten digits using fully connected Artificial Neural Network in FPGA which will act like a replacement for GPU

Abstract

AI/ML has grown significantly in the last few years, and slowly and steadily it is starting to impact the semiconductor industry as well. The world is moving towards more application specific technologies. We first saw CPU s, then GPUs came along for image specific applications, now, a new revolution is trying to take place by implementing Neural Networks in FPGA. In this paper, a 5-layer fully connected artificial neural network has been built using the MNIST dataset to recognize handwritten digits. Software like Vivado 2022.1 and in built Zynq 7000 board has been used to realize the implementation. An accuracy of 98% has been achieved through this implementation, and very minimal power and memory usage has also been seen. Regarding the paper, it has been clearly seen that ANN based implementation of FPGA is far superior to the workings of a GPU because of its efficient memory and power usage, also it creates less latency issue as FPGA works in a clock cycle of nanoseconds.