ASCENT Theme 4 Liaison Meeting -
Time: 4:00 pm - 5:00 pm
Title: Energy-Efficient Inference Using Fully-Integrated Reconfigurable CMOS-RRAM Compute-In-Memory Hardware
Presenter: Weier Wan (Prof. H.-S Philip Wong's Research group, Stanford)
Abstract: The Compute-In-Memory (CIM) architecture improves AI inference energy-efficiency by performing computation directly at where AI model is stored, eliminating expensive memory access for model weights. Resistive random-access memory (RRAM) enables CIM to scale towards larger model size by providing dense and analog on-chip weight storage. However, the greater energy-efficiency of CIM usually comes at the cost of reconfigurability, hampering its application for diverse AI workloads. The energy-efficiency itself is also often limited by the output sensing circuits. In this talk, we present the first fully-integrated RRAM-CMOS CIM chip that simultaneously offers dataflow reconfigurability to support diverse AI model architectures, and delivers the record energy-efficiency of 148 TOPS/W among RRAM-CIM hardware. The key to energy efficiency is the use of a novel voltage-mode sensing scheme. We demonstrated multiple AI tasks on the chip including image recovery using Restricted Boltzmann Machine and image classifications using Multi-Layer Perceptron. The chip was presented at 2020 ISSCC and 2020 Symp. VLSI Technology. Demo video: https://youtu.be/b7ITxmfaLBk
About speaker: Weier Wan is pursuing a Ph.D. degree at Stanford University, Department of Electrical Engineering. His research centers around building energy-efficient Compute-In-Memory hardware using resistive memories for AI acceleration. His research efforts span the full-stack of AI system including algorithm-hardware co-optimization, architecture and circuit design, chip tape-out and testing, and device characterization and modeling. Previously he received his Bachelor degree from University of California, Berkeley.
This meeting is only available to the JUMP research community, such as Principal Investigators, Postdoc researchers, Students, and Industry/Government liaisons.