Chenxu Niu

Chenxu Niu

HPC System Engineer @ NVIDIA

San Jose, CA NVIDIA

🔬 Research

My work spans three connected themes: measuring and reducing the energy cost of AI on HPC systems, applying LLMs to hardware design, and making scientific data discoverable.

Energy-Efficient AI & HPC Systems

Overview figure from TokenPowerBench, a benchmark for the power consumption of LLM inference

How much energy does serving a large language model really take, and where can it be saved? This line of work measures the power cost of LLM inference across GPUs, nodes, and serving engines, screens energy-aware serving configurations before they reach expensive accelerators, and builds power-centric observability for production HPC clusters.

Publications

Projects

View all 4 publications

LLMs for Hardware Design

Overview figure from FIXME, an end-to-end benchmark for LLM-aided design verification

Large language models are starting to help design the chips they run on. This work benchmarks LLM-aided design verification end to end and uses knowledge-guided LLM sampling to explore microarchitecture design spaces more efficiently than conventional search.

Publications

View all 2 publications

Scientific Data Management

Overview figure from ICEAGE, an LLM-powered engine for scientific data discovery

Scientific datasets are only useful if researchers can find what is inside them. This work makes self-describing scientific files searchable: learned representations of metadata attributes, parallel semantic querying, and LLM-powered exploration and answer generation for scientific data discovery.

Publications

View all 4 publications