Close

Presentation

Tackling Resource Utilization in DNN Accelerators
TimeTuesday, July 12th6pm - 7pm PDT
LocationLevel 2 Lobby
Event Type
Networking Reception
Work-in-Progress Poster
DescriptionThe method of dividing accelerator resources across multiple sub-accelerators is one of the currently used approaches to improve resource utilization in DNN accelerators. However, to ensure good performance across sub-accelerator systems, one must design a scheduler that efficiently maps DNN layers onto the system. We propose a scheduler that extends state-of-the-art to map layers on sub-accelerators systems in a manner that minimizes the energy-delay product with an average performance improvement of ~11.6%. Furthermore, we implement a Bayesian Optimization framework to automate the design of dataflows that would lead to the best performance for each sub-accelerator.