Performance Benchmarking of Scientific AI Workloads for Next-Generation HPC Systems

Project description

Scientific AI workloads are growing rapidly in scale and complexity, especially with the rise of foundation models and large-scale inference/training pipelines. 

This internship project will benchmark and analyze the performance of representative scientific AI workloads (such as materials characterization or weather forecasting) on NERSC systems, with emphasis on Perlmutter and relevance to future Doudna-class platforms. 

Project tasks

The intern will evaluate end-to-end performance across compute, communication, and I/O; study scaling behavior under different parallelism strategies; and profile bottlenecks in both training and inference settings.

Candidate workloads may include science-focused models and service-oriented stacks (e.g., vLLM-based inference), selected in collaboration with mentors and aligned with DOE/NERSC priorities, including the Genesis Mission.

The project outcome will be actionable benchmark results and analysis that help guide workload readiness, optimization priorities, and system design decisions for next-generation supercomputing.

Desired skills/background

Required

• Python
• Experience with ML workflows and/or performance analysis
• Software engineering fundamentals

Nice to have

• Distributed training/inference
• Profiling tools
• GPU communication and I/O optimization
• Familiarity with HPC environments

Back to 2026 projects