CAAM: Compressor-Based Adaptive Approximate Multiplier for Neural Network Applications

Abstract

The rapid growth of artificial intelligence and neural network applications has increased the demand for energy-efficient hardware accelerators. Multipliers are among the most resource-intensive components in neural network processors, consuming significant power and area. This project proposes a Compressor-Based Adaptive Approximate Multiplier (CAAM) that improves computational efficiency by utilizing adaptive error-tolerant compressors. The architecture reduces switching activity and critical path delay while maintaining acceptable computational accuracy. Approximate computing techniques enable a flexible trade-off between power consumption and output precision. The proposed design significantly lowers hardware complexity and area utilization compared to exact multipliers. Experimental evaluation demonstrates considerable energy savings with minimal degradation in neural network accuracy. The adaptive approximation mechanism allows scalability across different workloads and precision requirements. The design is modeled and verified using SystemVerilog to ensure reliability and functionality. CAAM provides an efficient and practical solution for next-generation edge AI and neural network accelerator systems.