NOVEL TUNABLE APPROXIMATE COMPRESSOR DESIGNS FOR ENERGY CONSTRAINED MULTIPLY-ACCUMULATE OPERATIONS IN EDGE AI

Abstract

Energy-constrained or self-powered hardware platforms are becoming more and more important for edge artificial intelligence (AI) systems, which require extremely efficient compute units without sacrificing functional precision. Modern AI workloads primarily use multiply-and-accumulate (MAC) processes, which means optimizing them is essential for sustainable edge intelligence. In this work, new tunable approximation compressor designs are presented that allow MAC topologies to dynamically change accuracy-energy trade-offs. The suggested compressors include tenability knobs and low-complexity approximation techniques that enable real-time configuration according to workload precision requirements or available energy budgets. When included into MAC units, the designs maintain a satisfactory level of accuracy for AI applications that require error resilience while achieving significant reductions in power consumption, delay, and silicon area. The designs are proven to be efficient and feasible through hardware implementation and assessment using Xilinx Vivado, proving their applicability for energy-adaptive computation in next-generation edge AI and autonomous sensing systems.