Approximate Arithmetic Circuits Enabling Energy-Efficient Edge Computing

Abstract

Approximate Computing (AxC) has emerged as a powerful technique for enabling efficient AI/ML workloads on resourceconstrained edge devices. By judiciously introducing controlled inaccuracies, AxC leverages the inherent resilience of AI/ML models to unlock substantial gains in power, performance, and area (PPA). Nonetheless, the widespread adoption of AxC is hampered by the need for applicationspecific designs, which increase design and implementation complexity. To address this, we highlight advanced techniques for approximate arithmetic operator design-focusing on LUT-level optimizations, automated operator modeling, and design-space exploration on FPGAs. Through these methods, AI-enabled Electronic Design Automation (EDA) can effectively reduce development complexity and drive large-scale AxC adoption, paving the way for energy-efficient, scalable AI computing at the edge.