Energy-Efficient Computer Systems: RISC-V Extensions for Machine Learning Inference at IoT's Edge Computing

Abstract

Over the past few decades, there have been numerous turning points in the massive transformation of computing systems. The limits of instruction-level parallelism (ILP) and the end of Dennard's scaling pushed the semiconductor sector toward multi-core devices, notwithstanding Moore's law, which directed the industry to pack more and more transistors and logic into the exact same volume. The era of domain-specific architectures (DSA) and processors for novel workloads like machine learning (ML) and artificial intelligence (AI) has recently begun. In addition to the difficulties brought on by tighter integration, extreme form factors, and increasingly varied workloads, these trends—possibly with additional limitations—further complicate the architecture, design, implementation, and power consumption optimization of systems. Nowadays, across the board, energy efficiency is a first-order design constraint and parameter for computing equipment. The creation of energy-efficient computer systems using RISC-V architecture modifications specifically suited for machine learning inference is investigated in this study. While preserving inference speed and accuracy, the suggested extensions seek to decrease energy consumption, minimize instruction overhead, and maximize hardware utilization. Through the integration of domain-specific accelerators, memory-access optimizations, and lightweight vector operations, the extended RISC-V platform exhibits notable gains in performance per watt when compared to traditional architectures. Results from experiments and benchmark assessments demonstrate how well-suited the method is for real-time Internet of Things applications including industrial automation, smart healthcare, and environmental monitoring. AI-enabled IoT systems that are low-power, scalable, and sustainable are being advanced by this work.