Designing Computer Systems under Sanctions and Sustainability

August Ning, Doctoral Assistant

Princeton University

Host: Babak Falsafi, PARSA

Recently events such as global chip shortages, semiconductor subsidies, and the rise of large scale machine learning have emphasized the importance of semiconductors in our everyday lives. Recent regulations by the United States of America placed export controls on flagship GPU devices used for AI applications. Governments and companies alike have policies toward reducing carbon emissions, specifically related to information and communications technology.

Computing sanctions and sustainability metrics fundamentally change how computer systems are designed and optimized. First, we will overview computing sanctions and how they affect chip architectures. Existing sanction specifications seem counterintuitive to how architects traditionally approach chip design. We show how existing policies make chip manufacturing more expensive and how LLM inference performance can still be improved compared to restricted baselines. Then I will detail how architecture-first policies can reduce the economic externalities caused by existing regulations.

Then we will focus on how computer system designers can reason about sustainability. We analyze fine-grained data center computing and energy data towards building a framework for evaluating sustainability and cost trade offs. To conclude, we will motivate the important role of computer scientists for crafting fair, effective, and impactful computing policy.

Bio:

August Ning is a senior PhD candidate at Princeton University, working with Prof. David Wentzlaff. His research interests span computer architecture under broadly defined economic constraints such as chip manufacturing, policy, and sustainability. He is supported by the NSF Graduate Research Fellowship and Princeton’s Gordon YS Wu Fellowship.