DATE 2021 became a virtual conference due to the worldwide COVID-19 pandemic (click here for more details)

Taking into consideration the continued erratic development of the worldwide COVID-19 pandemic and the accompanying restrictions of worldwide travelling as well as the safety and health of the DATE community, the Organizing Committees decided to host DATE 2021 as a virtual conference in early February 2021. Unfortunately, the current situation does not allow a face-to-face conference in Grenoble, France.

The Organizing Committees are working intensively to create a virtual conference that gives as much of a real conference atmosphere as possible.

IP6_4 Interactive Presentations

Date: Wednesday, 03 February 2021
Time: 17:00 - 17:30

Interactive Presentations run simultaneously during a 30-minute slot. Additionally, each IP paper is briefly introduced in a one-minute presentation in a corresponding regular session

Label Presentation Title
Authors
IP6_4.1 A CASE FOR EMERGING MEMORIES IN DNN ACCELERATORS
Speaker:
Avilash Mukherjee, University of British Columbia, CA
Authors:
Avilash Mukherjee1, Kumar Saurav2, Prashant Nair1, Sudip Shekhar1 and Mieszko Lis1
1University of British Columbia, CA; 2QUALCOMM INDIA, IN
Abstract
The popularity of Deep Neural Networks (DNNs) has led to many DNN accelerator architectures, which typically focus on the on-chip storage and computation costs. However, much of the energy is spent on accesses to off-chip DRAM memory. While emerging resistive memory technologies such as MRAM, PCM, and RRAM can potentially reduce this energy component, they suffer from drawbacks such as low endurance that prevent them from being a DRAM replacement in DNN applications. In this paper, we examine how DNN accelerators can be designed to overcome these limitations and how emerging memories can be used for off-chip storage. We demonstrate that through (a) careful mapping of DNN computation to the accelerator and (b) a hybrid setup (both DRAM and an emerging memory), we can reduce inference energy over a DRAM-only design by a factor ranging from 1.12x on EfficientNetB7 to 6.3x on ResNet-50, while also increasing the endurance from 2 weeks to over a decade. As the energy benefits vary dramatically across DNN models, we also develop a simple analytical heuristic solely based on DNN model parameters that predicts the suitability of a given DNN for emerging-memory-based accelerators.