CAREER: Elastic Algorithm-Architecture Co-Design for Energy-Efficient and Scalable Artificial Intelligence via Dynamic Connectivity

Artificial Intelligence (AI) is essential in modern applications from smartphones to autonomous vehicles and data centers. However, the growing demand of increasingly large models and more computing resources is leading to a rapid surge in energy usage. If left unaddressed, this trend will result in significant energy waste, limiting the potential of AI technologies while creating substantial economic and environmental issues. This project aims to identify and reduce unnecessary computations and data movements to save energy by making AI computing more flexible and adaptable. The significance of this project lies in rethinking how AI hardware processes, stores, and moves data, while creating an energy-aware design approach that will be openly available. The project will integrate research activities with education initiatives to engage students through enriched curriculum and outreach programs. These educational and outreach activities will also increase participation among local communities. This project addresses the challenges of energy efficiency and scalability in large AI systems, such as serving large language models, through a co-design approach with three key objectives. The central approach is that AI models, though traditionally viewed as static, can dynamically connect essential components to form computational graphs, enabling elastic processing with comparable performance while reducing the costs of redundant components. The first objective focuses on developing methods to leverage dynamic connectivity to reduce redundancy by adapting models to specific tasks and data. The second objective involves providing architectural support for elastic processing by exploring heterogeneous architectures that integrate approximate, analog, and near-data computing. The third objective aims to enable hardware-aware model adaptation, through co-design space exploration of algorithms and architectures for greater efficiency. These objectives collectively seek to overcome the energy wall by reducing computational costs while maintaining model performance, advancing future paradigms for energy-efficient and scalable AI systems. 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.