Electrical engineering and computer science Professor Pengfei Su has received a CAREER award for his research on cross-layer performance tuning to enhance deep learning model efficiency.
He is the 41st researcher from UC Merced to earn a CAREER award from the National Science Foundation (NSF).
CAREER awards are among the NSF's most prestigious awards. They are given through the Faculty Early Career Development Program to recognize untenured faculty members as teacher-scholars. Early-career faculty members are selected based on three factors: the strength of their research proposals; their potential to serve as academic role models in research and education; and their leadership in their field and organizations.
Su will receive $604,250 over the next five years for the project "Reforming Profiling Techniques to Guide Systemic Performance Tuning for GPU-Accelerated Deep Learning Workloads."
The project aims to establish a systematic approach for tuning deep learning models. It is structured around three techniques: unified binary code analysis to identify inefficient code segments and data objects across layers, incremental analysis to refine performance monitoring to locate the root causes of inefficient code segments and data object analysis to diagnose the root causes of inefficient data objects.
Su and his lab specialize in programming languages, program analysis, high-performance computing, machine learning systems and software engineering, with a focus on developing tools to analyze and optimize software inefficiencies.
"Graphics processing units (GPUs) are the powerhouse of deep learning, delivering unmatched computational performance," Su said. "Yet, fully unlocking their potential remains a formidable challenge as deep learning models grow increasingly complex, spanning multiple abstraction layers.
"While this complexity fuels innovation across diverse applications, it also introduces hidden inefficiencies that arise from intricate cross-layer interactions," Su said. "This project will pioneer a comprehensive, cross-layer performance analysis to expose these inefficiencies, optimize execution and push GPU performance to new frontiers."
Deep learning uses artificial neural networks to teach computers to learn and make decisions. It is a subset of artificial intelligence often used to recognize complex patterns in data, such as images, text and sounds.
Su has been with UC Merced since 2021. He is affiliated with the High Performance Computing Systems and Architecture Group.
Each CAREER award proposal includes educational outreach. The project's findings will be integrated into computer science curricula and K-12 programs to cultivate a new generation of experts in performance analysis and optimization, ensuring lasting contributions to academia and industry.
Su was honored to learn about the CAREER award.
"This award represents a major step forward in fundamental software performance analysis, driving a deeper understanding of performance challenges in deep learning," Su said. "By strengthening the foundations of program analysis, the project will accelerate breakthroughs in many AI-driven fields, such as image processing. Additionally, with interest from industry leaders like Meta, the project is poised to translate cutting-edge research into real-world impact."
