Zhaoxin Ban

About Zhaoxin Ban

I am interested in understanding the behavior of large machine learning systems.My current work focuses on performance analysis of ML inference workloads. I examine latency patterns, GPU utilization, kernel execution behavior, and memory bandwidth usage to understand why a model runs slowly. Using tools such as Nsight Systems and Nsight Compute, I analyze execution traces and hardware counters to distinguish compute-bound and memory-bound behavior and to study how computation and memory access overlap during model execution.I also look at how compilation and kernel granularity influence runtime efficiency. This includes comparing compiled and non-compiled execution paths, understanding kernel launch overhead, and observing how kernel size and fusion affect hardware utilization.Rather than treating performance tuning as parameter tweaking, I approach it as a system diagnosis problem. The internal execution state of a model cannot be directly inspected; only indirect signals such as timing profiles, utilization metrics, and scaling behavior are visible. From these signals, I try to infer waiting sources, execution inefficiencies, and resource contention, and I often build small diagnostic or validation tools (e.g, offline consistency checks) to verify hypotheses about execution behavior and correctness.This way of thinking comes from my background in computational hydrology. Hydrologic systems are also only partially observable: we observe precipitation and discharge, but not the internal storage and transport processes. My research involved inferring hidden system dynamics from sparse observations — essentially an inverse problem and system identification task. I find the same reasoning applicable when analyzing modern ML systems.More broadly, I am interested in:• ML inference efficiency• GPU execution behavior and roofline characteristics• performance debugging and attribution• correctness and reproducibility of model executionOutside of work, I also compose music and write long-form fiction — another form of modeling sequential processes

Experience

1. Machine Learning Engineer

   ByteDance

   Jun 2025 — Present · Bellevue, WA, US

Education

• Nanjing University

  Bachelor of Science - BS, Geography and Hydrology

• UCLA

  Doctor of Philosophy - PhD, Computational Hydrology

• UCLA

  Master's degree, Computational Hydrology