← All episodes

Ep 17: Vijay Reddi | Architecture 2.0 and AI for Computer Systems Design

Listen on Podbean →

Summary

In this insightful episode of the Computer Architecture Podcast, hosts Dr. Suvinay Subramanian and Dr. Lisa Hsu are joined by Vijay Janapa Reddi, an Associate Professor at Harvard University and Vice President and Co-founder of ML Commons. Professor Reddi is a distinguished figure in the field, known for his substantial contributions to mobile and edge computing systems, his key role in developing the MLPerf benchmarks, and his passion for education, exemplified by his machine learning systems book, `mlsysbook.ai`.

The conversation delves deep into the concept of "Architecture 2.0," a new paradigm shift where AI and Machine Learning are not just applications but integral tools for designing future computer systems. Professor Reddi shares his vision for this new era, discussing the exciting opportunities it presents, the inherent challenges, and the necessary educational shifts to equip the next generation of architects. He explains how Architecture 2.0 fundamentally rethinks the use of AI/ML to build more complex and efficient systems, especially in an age of generative AI.

A significant portion of the discussion also focuses on Professor Reddi's work in "TinyML." This fascinating area explores how to enable machine learning on resource-constrained devices, such as those with only hundreds of kilobytes of memory and milliwatt-level power consumption. He elaborates on the unique challenges and the immense potential of TinyML to transform our technological interactions, making AI truly ubiquitous by embedding it into everyday objects, from coin-cell battery-operated devices to always-on components in smartphones. The episode also touches upon the crucial role of community, open-source contributions, and collaborative efforts in driving these advancements forward.

Professor Reddi’s personal journey and inspirations, including his unique take on classic computer architecture textbooks and the importance of mentorship, add a personal and engaging dimension to the technical discussion, making this episode a compelling listen for anyone interested in the future of computer architecture and AI.

Chapters

• 00:01:32 — Introduction and What Motivates Vijay Janapa Reddi
• 00:02:50 — Defining Architecture 2.0: AI/ML for System Design
• 00:05:16 — The Role of Data in Architecture 2.0: Beyond ImageNet
• 00:08:14 — Pillars of Data for Architecture 2.0: Generation, Optimization, and Prediction
• 00:11:48 — Breaking Abstraction Layers: The Horizontal Design Space
• 00:15:48 — Data Collection and Quality for Architecture Research
• 00:25:29 — The Importance of Community and Open Source in Architecture 2.0
• 00:34:36 — TinyML: Enabling ML on Resource-Constrained Devices
• 00:43:05 — Educational Resources for TinyML and Modern Architecture
• 00:51:07 — Vijay Janapa Reddi's Journey into Computer Architecture and Academia
• 00:57:02 — The Future of Collaborative Research and Generational Shifts
• 00:58:54 — Concluding Thoughts and Episode Wrap-up

Takeaways

• Architecture 2.0 signifies a paradigm shift where AI and ML are used not just as applications but as fundamental tools to design and optimize future, increasingly complex computer systems.
• Data is the new code in Architecture 2.0; creating high-quality, domain-specific datasets (akin to an "ImageNet for computer architects") is crucial for training AI models to assist in system design.
• The concept of a horizontal design space in Architecture 2.0 involves breaking traditional abstraction layers and exposing parameters across these layers, allowing AI agents to holistically optimize system design.
• TinyML focuses on enabling machine learning on highly resource-constrained devices (kilobytes of memory, milliwatts of power), requiring a co-design approach involving hardware, software, and ML algorithms to make AI truly ubiquitous.
• Community collaboration and open-source efforts (like MLPerf and open educational resources) are vital for bootstrapping new architectural paradigms like Architecture 2.0 and TinyML, fostering shared knowledge, benchmarks, and tools.