Collaborative Research: SHF: Small: HexAI: Holistic Exploration for Design Automation of Efficient AI Accelerators

Rutgers University New Brunswick, New Brunswick NJ

Investigators

Abstract

Artificial intelligence (AI) systems are rapidly transforming a wide range of industries, from healthcare to manufacturing, by enabling machines to perform complex tasks such as pattern recognition, natural language understanding, and autonomous decision-making. These systems rely on AI hardware accelerators for efficient computing, but designing such accelerators remains a time-consuming, manual process that limits innovation and accessibility. This project addresses the national need for scalable and energy-efficient AI infrastructure by automating the design of AI accelerators. The new framework developed in this project aims to significantly reduce the time and effort needed to produce high-performance, low-power AI hardware. By accelerating the deployment of efficient AI systems, this work can help make advanced AI capabilities more broadly accessible and sustainable, fostering economic growth and scientific progress across domains. This project develops HexAI, a comprehensive design automation framework for next-generation AI accelerators. HexAI integrates architectural optimization, workload and data-awareness, and backend circuit synthesis into a unified framework. It introduces three major innovations: (1) a structured design space representation that supports efficient exploration of architectural and mapping decisions; (2) data- and workload-aware design strategies that adapt hardware to real-world AI tasks, including training and multi-tenant inference; and (3) backend-aware optimization using predictive register-transfer level (RTL) modeling and multi-fidelity Bayesian optimization methods to jointly improve power, area, and timing characteristics. By jointly addressing design space complexity, real-world adaptability, and RTL-level implementation challenges, HexAI aims to achieve an order-of-magnitude improvement in energy efficiency and significant reductions in design time. The resulting tools and insights will advance the field of electronic design automation and enable fast deployment of energy-efficient AI hardware. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.