Session 29 Overview: ML Chips for Emerging Applications

Summary form only given, as follows. The complete presentation was not made available for publication as part of the conference proceedings. Beyond conventional ML models, emerging ML algorithms and applications evolve at a fast pace, such as recommender systems, transformers, and spiking neural networks. Such emerging ML models exhibit different model architectures and computation/storage requirements, necessitating innovations in custom hardware design. This session includes four papers, each representing advances in ML chips for emerging applications with new techniques, such as 3D hybrid bonding, approximate/sparse computing, reconfigurable digital computing-in-memory, and spike-based on-chip learning.