Flyer: Integrated Data-Centric Computing for Efficent AI

The increasingly more complex artificial intelligence models have attracted interdisciplinary efforts to speed up AI workloads. The disruptive paradigm of data-centric computing is a promising approach to overcome the memory wall problem that conventional platforms face. Data-centric computing aims to perform processing closer to where data is stored or generated, shorten the distance of data movement, and reduce data dimensionality as much as possible. Also, advanced heterogeneous integration techniques are crucial to further assemble diverse hardware components into a compact form factor for enhanced energy and area efficiencies. This talk focuses on the circuit design and heterogeneous integration aspects of data-centric computing. Circuit implementations of in-memory computing and on-sensor computing engines will be discussed. The unique opportunities to transform computing with 2.5D and 3D integration technologies will be presented. Finally, case studies on how these techniques come together for system-level optimizations will be shared.