Close

Presentation

Wear Leveling of Processing Elements Array in Deep Neural Network Accelerators
DescriptionThe primary focus of DNN optimizers is on finding performance or energy optimal mapping while oblivious to reliability of DNN accelerator. The resulting DNN mapping utilizes only a limited part of PEs and starting point of mapping space is biased to the upper-left corner of PE array. As a result, biased-stress to specific PEs leads to reliability problem. To balance the PE usage and resolve the reliability problem, this paper proposes: i) torus-connected PE array that enables relocation of mapping space, ii) wear leveling technique that mitigates the unbalanced PE usage. Consequently, proposed technique improves the lifetime up to 3.9x.
Event Type
Work-in-Progress Poster
TimeTuesday, July 11th6:00pm - 7:00pm PDT
LocationLevel 2 Lobby
Topics
AI
Autonomous Systems
Cloud
Design
EDA
Embedded Systems
RISC-V
Security