IEEE HART: Every Bit Matters - Andreas Moshovos

Advances in computing hardware and system software have been a primary enabler of machine learning’s rapid progress and its impact across science, industry, and the global economy. Further gains are increasingly limited by fundamental barriers in power efficiency, memory capacity, and data movement costs. This talk will highlight some of these emerging challenges in computing for machine learning. It will also present a hardware/software co-design perspective on allowing further improvements. We will highlight recent work on datatype learning—the automated discovery and deployment of optimal numeric representations—and on techniques for optimized data encoding that minimize transfer overheads while preserving accuracy. Together, these approaches chart a path toward substantially more efficient deep learning training and inference, enabling continued scaling of model capability under practical constraints.

Presented by Andreas Moshovos, University of Toronto, ByteShape Inc.