Approximate Softmax Architecture for Energy-Efficient Deep Neural Networks

Read full text Summarise with VIRAT
Authors: Athul Krishna R, Deepak S J, Fayas Mohamed A, Baskar S
Cited by: 0
References: 7

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Softmax is commonly used in neural network- based classification systems to convert output values into probabilities, but its conventional implementation involves complex operations such as exponentials and division, which are inefficient for FPGA and VLSI-based hardware due to high area, power, and latency requirements. This project presents a hardware- efficient approximate softmax architecture designed for low-power FPGA systems using a simplified Top-1 approximation approach. The proposed design identifies the dominant output class using comparator-based logic and fixed- point arithmetic, thereby eliminating computationally expensive operations while maintaining correct decision-making. The approximate softmax module is implemented using synthesizable SystemVerilog and validated through simulation and synthesis using Xilinx Vivado, with Python used only for basic numerical verification. The developed design is hardware-ready and suitable for integration into FPGA-based neural network accelerators and embedded VLSI systems, with scope for future hardware implementation.

Fields of Study & Subjects

Low-power high-performance VLSI design Advanced Neural Network Applications Numerical Methods and Algorithms

Publication details

Journal / source
International Journal of Science Strategic Management and Technology
Volume
02
Issue
03
Pages
1–9
Publication date
25 March 2026
Language
EN