Evaluating HLS4ML For Converting a CNN Model Into High Level Synthesis

Abstract

Deep learning models, although high-performing, often require hardware acceleration to be effectively deployed on Field Programmable Gate Arrays (FPGAs). This paper investigates the use of the hls4ml framework to convert a Convolutional Neural Network (CNN) model into High Level Synthesis (HLS). hls4ml is a Python tool that translates neural networks into FPGA-compatible firmware for high-performance inference with reduced latency and power usage. It employs High-Level Synthesis (HLS) techniques to translate high-level programming languages like C++ or Python into hardware description languages, such as VHDL or Verilog. We evaluate the model’s performance in terms of accuracy and demonstrate the effectiveness of converting a deep learning model, specifically a CNN, for hardware implementation. The results highlight the potential of hls4ml facilitating embedded artificial intelligence applications, offering an efficient and scalable solution for deploying deep learning models on resource-constrained platforms.