Argonne extreme-scale computing workshop provides hands-on experience with supercomputers and AI

A buzz filled the auditorium as students and researchers from academia, government and industry began a session about vibe coding at the Argonne Training Program on Extreme-Scale Computing (ATPESC), an annual program hosted by the U.S. Department of Energy’s (DOE) Argonne National Laboratory and its Argonne Leadership Computing Facility (ALCF). Participants put new artificial intelligence (AI)-assisted approaches to coding into practice using their laptops to tap into the power of DOE supercomputers.

That energy carried throughout the intensive two-week program, which provides training on key skills, approaches and tools to do research on exascale and next-generation computing systems. This year’s program was held from July 26 to Aug. 7 at the Q Center in St. Charles, Illinois. It offered various tracks, including expanded topics on AI and machine learning (ML), as well as vibe coding, a technique that uses natural language prompts to guide AI to write or modify computer code instead of writing it manually line by line.

This year, 75 participants took part in ATPESC, bringing the program’s total attendance to nearly 1,000 people from 47 institutions since its launch in 2013. ATPESC offers sessions with leading experts in high performance computing (HPC) and allows participants to use some of the world’s most powerful supercomputers: Aurora and Polaris at the ALCF and Frontier at the Oak Ridge Leadership Computing Facility (OLCF) at DOE’s Oak Ridge National Laboratory. The ALCF and OLCF are DOE Office of Science user facilities.

“The hands-on sessions and direct access to experts are what make this training so special and so valuable. I know of no other program that commits to that level of mentoring.” — Bruno Lopez, Metacomp Technologies senior research scientist

Bruno Lopez, a senior research scientist at Metacomp Technologies in Westlake Village, California, primarily used Aurora and Polaris during his stint with ATPESC.

“The hands-on sessions and direct access to experts are what make this training so special and so valuable,” Lopez said. ​“I know of no other program that commits to that level of mentoring. Being able to log into some of the fastest computers in the world, experiment with the various HPC tools, submit a job and iterate on it, applying immediately what has just been presented, is something no lecture can substitute for.”

For Lopez, the combination of resources, people and format were valuable.

“The resources, because you are given direct access to leadership-class machines and to the whole ecosystem of tools that surrounds them,” Lopez said. ​“The people, because the presenters are the very experts who built what you are using, people with a real passion for their subject and a genuine desire to pass on what they know. And the format, because two intensive weeks is long enough to go deep into each subject.”

Such access allows participants to get hands-on experience on how the supercomputers work and how they drill down for answers, according to Deborah Stevens, an Argonne assistant computational scientist and member of the ATPESC organizing committee.

“Using the supercomputers has been valuable for participants, including researchers who work in fields other than computer science,” Stevens said. ​“These sessions teach them how to improve performance on the big machines and, maybe, how to build even bigger machines.”

Sharanjeet Kaur, a postdoc appointee in Argonne’s Transportation and Power Systems division, explored the architectures of different exascale computers, programming models, visualization tools, HPC debugging and profiling tools, agentic AI workflows, quantum computing and best practices in HPC software development. She plans to apply the program’s lessons directly to her work.

“In particular, I want to use the in situ visualization approaches I learned in the visualization track for my simulation cases,” Kaur said. ​“I also plan to make greater use of debugging and profiling tools to improve code performance and reduce manual debugging effort. In addition, I am highly motivated to adopt the concept of lab notebooks to better document numerical experiments and to incorporate the software development best practices. Finally, I plan to continue exploring and developing agentic AI workflows as part of my future work.”

Jamie Bamber, a postdoctoral research associate at the University of Illinois Urbana-Champaign, said that in his field of numerical relativity, he had always heard people talk about ATPESC in ​“glowing terms,” even in the United Kingdom, and ​“I can now very much echo that endorsement.”

“One of my key goals in going to ATPESC was to learn how to develop efficient software for modern GPU-based supercomputers, such as those hosted at Argonne,” Bamber said. ​“I have already started applying some of those techniques to our current project and hope to run exascale black hole simulations on Aurora and other machines soon. I am also keen to explore how we can use AI to accelerate our research in the field of numerical relativity, using some of the techniques presented at ATPESC.”

Michael A. Boateng, a Ph.D. student at Georgia Tech in the AI Institute for Advances in Optimization lab, with a joint appointment with DOE’s Los Alamos National Laboratory, also worked on these supercomputers during ATPESC.

“For me, ATPESC is unique because it combines technical depth, hands-on training and community,” Boateng said. ​“We were not just learning about HPC concepts theoretically. We could work with examples and experiment with different configurations. For example, the Portable, Extensible Toolkit for Scientific Computation software library training let us change nonlinear solvers and preconditioners and observe their effects, while the tools track allowed participants to work with performance and debugging tools and apply them to their own applications.”

Boateng noted that learning just wasn’t confined to a supercomputer.

“The community was also a major part of the experience,” Boateng said. ​“I met people from the United States, Italy, Spain, India, Brazil, France, Belgium, Ghana, China and other countries. We went out for sushi, deep dish pizza, visited the Skydeck and the Bean, went to museums and even tried indoor rock climbing! Those experiences made ATPESC feel like more than a training program. The two-week format gave us time to learn from one another and build friendships outside of the classroom as well.”

Before Amal Gueroudji became an assistant computer scientist in Argonne’s Mathematics and Computer Science division, she attended ATPESC in 2021 and returned this year as an instructor. In 2021, she was a second-year doctoral student in France, and her advisor started pushing her to think seriously about what she wanted to do next. He suggested that she consider Argonne.

She remembered telling him, somewhat jokingly, ​“If I go to the U.S., I want to work for NASA.”

Her advisor’s response was, ​“Argonne is the NASA of supercomputing.”

Participating in ATPESC gave her the first real window into what he meant, Gueroudji said.

“Through the program, I was exposed not only to the breadth of HPC research, but also to the people, systems and scientific challenges at Argonne and across the DOE computing community,” she said. ​“It made Argonne feel like a place where I could see myself doing research. Not long afterward, I applied for and obtained a postdoctoral position at Argonne, and today I am a staff scientist here. So, in a very real way, ATPESC influenced the trajectory of my career. Coming back five years later as an ATPESC instructor makes that journey particularly special.”

This year’s ATPESC expanded instruction on AI/ML, and participants worked with the ACLF AI Testbed machines. They also accessed the ALCF Inference Service for research and vibe coding, said Raymond Loy, director of ATPESC and ALCF lead for training, debuggers and math libraries.

This year, the ATPESC committee expanded the AI/ML track from one day to two, covering topics such as distributed deep learning, introduction to large language models, pretraining, post training, workflows, inference and agentic tools. In addition to the expanded AI/ML track, they also added AI-adjacent topics in mixed-precision computing and vibe coding, Loy said.

The emergence of DOE’s Genesis Mission also influenced the direction of how ATPESC can best train the future generation of HPC professionals. The historic Genesis Mission will deliver breakthroughs using AI to secure American energy dominance, accelerate scientific discovery and strengthen national security.

“ATPESC strives to be at the forefront of computational science, and the curriculum is adjusted every year,” Loy said. ​“AI and ML have become increasingly more important for science — as witnessed by the Genesis Mission — and we have followed suit.”