Cheng-Hsiu Hsieh

Resolved TSMC's 2nm legalizer pin access challenges by reformulating geometric verification as parallel string matching on GPU. Achieved 25x+ performance improvement through CUDA kernel optimization with shared memory and reduced bank conflicts. Integrated solution into Fusion Compiler backend, impacting billion-scale verification API calls in production.

Developed and deployed time series forecasting models (XGBoost, CNN+LSTM) on AWS EC2 using a 2-year market dataset. Ranked 1st (out of 20 interns) in the firm-wide competition, outperforming the existing model by 3.3% in forecast accuracy.

Engineered an out-of-band C++ daemon to monitor power faults on 75k+ production x86 servers via I2C and PMBus protocols, reducing troubleshooting time by 150x and slashing maintenance costs by over 90% through automated detection and analysis. Created extensible diagnostic framework integrating with OpenBMC ecosystem through D-Bus messaging services and Redfish API endpoints, designed for multi-vendor compatibility using JSON-driven configuration architecture. Drove collaboration between hardware vendors and internal power teams to establish unified fault reporting standards, shipping production solution with comprehensive testing and technical onboarding documentation.

Introduction to Generative Artificial Intelligence @ NTUEE

Finalist in WorldQuant BRAIN Alphathon

Algorithm @ NTUEE