Schedule

08:00 - 08:10 am Opening remarks

Presenter: Prof. Yingyan (Celine) Lin

08:10 - 08:55 am Introduction and background knowledge

Presenter: Yongan Zhang

• Deep Neural Network (DNN) workloads

  • DNN basics

  • Why are DNNs prospective?

  • Why do we need efficient models and hardware accelerators?

  • Recent progress in efficient DNN models and accelerators

• Challenges in developing efficient DNN models and accelerators

• A promising solution → Co-search and thus this tutorial

08:55 - 09:45 am Auto-NBA: Co-search algorithm and network design space

Presenter: Yonggan Fu

• Background:

  • • What is the role of hardware search space and modeling?

• Motivation:

  • • Why is it crucial?

• Efficacy:

  • • What does it enable and what are the use cases?

• Technical Details

• Examples+Guidance:

  • • Toy example illustration

09:45 - 10:00 am Auto-NBA: Hardware design space and performance modeling

Presenter: Yongan Zhang

• Role of hardware subpart in Auto-NBA

• Use cases beyond Auto-NBA

• Detailed methodology

10:00 - 10:20 am HW-NAS-Bench (1)

Presenter: Chaojian Li

• Challenges in Co-Search When Target More (Commercial) Devices

• Highlighted Features of Our Proposed HW-NAS-Bench

• Analysis offered in HW-NAS-Bench

• Real Use Cases Reported by Other Researchers

10:20 - 11:00 am Q&A + Coffee break

11:00 - 11:25 am HW-NAS-Bench (2)

Presenter: Chaojian Li

• Hands-on Demonstration

11:25 - 11:55 am DNN-Chip Predictor

Presenter: Yang Zhao

• Background:

  • • DNN-Chip Predictor's Role in Co-Search

• Overview:

  • • What You Can Do with DNN-Chip Predictor

• Technical Details:

  • • DNN-Chip Predictor's Analytical Models

• Evaluation:

  • • Evaluating DNN-Chip Predictor

• Guidance:

  • • How to Use DNN-Chip Predictor

11:55 - 12:00 pm Q&A