ChatHPC: Building the Foundations for a Productive and Trustworthy AI-Assisted HPC Ecosystem
Date: January 20th, 2026
Time & Location: Pettit Microelectronics Research Center (MIRC) 102B, 11 AM – 1 PM
The Center for Scientific Software Engineering invites you to this seminar and hands-on event from Oak Ridge National Lab researchers. The event will be held in-person, and faculty, staff, and students are welcome to attend. We ask that you register by close of business January 15th so that we can plan for lunch.
Register here
Presenters:
Pedro Valero Lera – Senior Computer Scientist; Programming Systems Group, Oak Ridge National Lab (ORNL)
Aaron Young – Software Engineer; Architectures and Performance Group, ORNL
Keita Teranishi – Senior Computer Scientist and Group Lead; Programming Systems Group. ORNL
Abstract:
ChatHPC aims to democratize the use of large language models (LLMs) within the high-performance computing (HPC) community by providing an accessible infrastructure and ecosystem for applying generative AI technologies to critical HPC components. Our approach combines technical innovation with practical workflows, enabling participants to rapidly create specialized AI assistants for HPC tasks using modest computational resources.
In this session, we will first present the core concepts behind ChatHPC, including our divide-and-conquer strategy for building reliable, optimized assistants through cost-effective fine-tuning of Meta’s Code Llama models under expert supervision. These assistants target key areas of the HPC software stack—programming models, runtimes, I/O, tooling, and math libraries—helping improve portability, parallelization, optimization, scalability, and instrumentation.
Following the technical overview, we will conduct a hands-on walkthrough demonstrating how small datasets (on the order of KB) and a single node with two NVIDIA H100 GPUs can fine-tune a 7-billion-parameter Code Llama model in minutes to deliver high-quality, trustworthy solutions. We will also compare performance and trustworthiness against larger models such as OpenAI’s GPT-4o, highlighting how ChatHPC achieves up to 90% higher reliability for critical programming tasks.
Join us to explore how ChatHPC can accelerate HPC software development and foster a more productive, AI-assisted ecosystem.