In this episode, hosts Dr. Suvinay Subramanian and Dr. Lisa Hsu welcome Dr. Vivienne Sze, an Associate Professor in the EECS department at MIT. Dr. Sze is renowned for her leading work in energy-efficient computing systems, with expertise spanning video compression, machine learning, robotics, and digital health. Her accolades include a DARPA Young Faculty Award, an Edgerton Faculty Award, and a Primetime Engineering Emmy for her role in developing the high-efficiency video codec (HEVC) standard.
The discussion delves into the intricate world of energy-efficient algorithm-hardware co-design, particularly for compute-intensive applications. Dr. Sze shares her insights from teaching a hardware for deep learning class at MIT, emphasizing the importance of distilling fundamental principles in rapidly evolving fields. The conversation explores the challenges and strategies in designing efficient systems across diverse domains, from the sequential dependencies in advanced video compression algorithms to the unique demands of deep neural networks and robotics. Dr. Sze also reflects on her journey, the importance of collaboration, and the often-overlooked non-technical skills crucial for success in technical fields.
Chapters
00:00:00 — Introduction to the Computer Architecture Podcast
00:00:13 — Introducing Guest Dr. Vivienne Sze and Her Work
00:01:19 — What Gets Vivienne Sze Up in the Morning: Teaching, Collaboration, and Societal Challenges
02:57 — Distilling Principles in Rapidly Evolving Fields: The Hardware for Deep Learning Class
03:42 — Generalizing Principles from Research for Broader Impact
05:13 — Trade-offs in Video Encoding: Parallelism vs. Work Efficiency
06:05 — Understanding Video Compression: Redundancy and Prediction
08:32 — Parallels and Differences: Video Compression vs. Deep Neural Networks (DNNs) for Energy Efficiency
09:52 — Comparing Video Coding and Deep Learning: Standardization vs. Flexibility in Hardware Design
13:40 — Balancing Flexibility, Performance, and Efficiency in DNN Compression
15:00 — Accuracy vs. Performance Trade-offs: Beyond Operations Count to Hardware Metrics
20:13 — Defining and Measuring Accuracy in Application-Specific Contexts: Robotics and Healthcare
25:25 — Understanding System-Level Trade-offs in Robotics and Healthcare
25:52 — End-to-End System Design: Pruning the Design Space Across Algorithms and Hardware
31:14 — The Carbon Footprint of Autonomous Vehicles: A Sustainability Perspective
35:32 — Dr. Vivienne Sze's Journey: From Undergrad Internship to MIT Faculty
37:58 — The Story Behind the Primetime Engineering Emmy for HEVC Standard
45:51 — The Societal Impact and Evolution of Video Compression Standards
48:02 — Dr. Vivienne Sze's Words of Wisdom: The Importance of Collaboration and Non-Technical Skills
54:22 — The Limiting Factor: Interpersonal Skills and Handling Feedback in Technical Careers
57:22 — Final Thoughts and Thanks to Dr. Vivienne Sze
57:50 — Closing Remarks
Takeaways
In rapidly evolving fields like AI hardware, it's crucial to distill fundamental principles rather than just chasing the latest trends, enabling more robust and lasting system designs.
Energy-efficient system design involves a deep understanding of trade-offs, such as parallelism for hardware efficiency versus the sequential dependencies introduced by advanced algorithms in video compression or the flexibility required by diverse DNN workloads.
Data movement is a critical bottleneck for energy efficiency in both video compression and deep neural networks; minimizing it through co-design is key.
Effective problem-solving in complex domains like robotics and healthcare requires strong interdisciplinary collaboration and a willingness to learn from experts in different fields to understand true application needs and constraints.
Non-technical skills, such as the ability to give and receive constructive feedback, manage conflict, and communicate effectively across disciplines, are as vital as technical expertise for success and impact in research and engineering.