Yifan Sun

Yifan Sun

(孙轶凡)

Class of 1953 Associate Professor

Computer Science

William & Mary

location
ISC 4 1383
540 Landrum Drive
Williamsburg, VA 23185
file CV
email profyifansun@gmail.com
google scholar Google Scholar
github GitHub
xiaohongshu 小红书
bilibili 哔哩哔哩

Yifan Sun

I am the Class of 1953 Associate Professor of Computer Science at William & Mary. I am the lead of HUSCARL (Human-Centered System & Computer Architecture Lab). I received my Ph.D. degree from the Department of Electrical and Computer Engineering at Northeastern University in 2020. My research interests lie in GPU architecture, performance evaluation, and performance modeling.

Join My Lab

If you are interested in joining HUSCARL, either as a Ph.D. student, master's student, undergraduate researcher, or research intern, you can contact me by sending an email to profyifansun@gmail.com.

Research Topics

Human-Centered Performance Analysis
Modern chips are too complex for designers to reason about from aggregate statistics. We build visualization tools and AI assistants, such as Daisen and DaisenBot, that show architects where time goes, why a bottleneck occurs, and what to change. We also study how experts actually perform performance analysis.
AI Agents for Computer Systems Research
AI agents can now write code, run experiments, and read results, but they still make mistakes that humans must catch. We design multi-agent systems, such as TheBotCompany, that develop and maintain research software, and we study how humans and agents can explore architecture designs together while humans stay able to verify the results.
Performance Modeling for Large-Scale AI Workloads
Cycle-level simulation cannot keep up with LLM training and inference that span thousands of GPUs. We develop lightweight, validated performance models, such as TrioSim, on top of our open-source Akita and MGPUSim frameworks, and we evaluate which modeling methods are accurate enough for which design questions.
Wafer-Scale and Multi-GPU Systems for AI
AI workloads are pushing GPUs beyond a single die, toward wafer-scale and multi-GPU systems where data movement dominates. We design address translation, memory, and interconnect techniques, including electro-photonic networks, that keep these systems efficient (e.g., HDPAT).