KEYWORD: ARCHITECTURE & SYSTEM DESIGN, ANALOG & MIXED SIGNAL, EMERGING TECHNOLOGIES
EVENT TYPE: PANEL
In Memory Computing for Machine Learning - Future or Fallacy?
Reetuparna Das - Univ. of Michigan, Ann Arbor, MI
Valeria Bertacco - Univ. of Michigan, Ann Arbor, MI
Several flavors of memory-centric computing has evolved over the years. Back from 90s were Processing-in-Memory or Near-Memory architectures were born in form of logic units tightly fused to separate memory structures, to present day In-Memory architectures which repurpose memory structures to do computing. Further, these class of architectures has the added dimension which memory technology to be used.
This panel will focus on the advantages and pitfalls of using memory-centric computing for AI. The champions are likely to make a case based on data-movement benefits, area savings, parallelism, integration benefits, and efficiency. The other side is likely to make an against case based on precision, flavors of analog in-memory computing which introduce data conversion overheads, technology feasibility, and programmability.
Daniel Morris - Facebook Rangarajan Venkatesan - NVIDIA Corp., Santa Clara, CA Nuwan Jayasena - Advanced Micro Devices, Inc., Santa Clara, CA Shekhar Borkar - Qualcomm Technologies, Inc., San Diego, CA Engin Ipek - Univ. of Rochester, Rochestor, NY