Introduction to AI Training Workflows for HPC Support Teams

This course takes place on-site at JSC. Online participation is not possible. The course will be held in English.

Course Content:

• Introduction to deep learning: key concepts of modern AI training workflows
• Basics of PyTorch and implementing an AI training workflow
• Setting up a virtual environment and running training workloads on HPC systems
• Monitoring training using TensorBoard and Weights & Biases
• Parallelisation strategies for distributed training
• Profiling of training code, identifying bottlenecks and improving performance

 

After the course, participants will be able to:

• Explain the basic concepts and terminology of deep learning
• Understand and adapt an AI training workflow in PyTorch
• Help users run AI training workflows on HPC systems using suitable parallelisation strategies
• Use NVIDIA Nsight Systems to identify performance bottlenecks in training and advise users on potential improvements

 

Training methods used in this course:

• Presentations introducing the concepts
• Hands-on sessions on the JURECA supercomputer, including running deep learning workloads, collecting profiling traces and analysing the results
• Q&A sessions and discussion

 

Contents level

in hours

in %

Beginner's contents:

3 h

21 %

Intermediate contents:

6 h

43 %

Advanced contents:

5 h

36 %

Community-targeted contents:

0 h

0 %

 

Prerequisites:

Participants should have experience with HPC systems and bring their own laptop.

A personal institutional email address (university/research institution, government agency, organisation, or company) is required to register for JSC training courses. If you don't have an institutional email address, please get in touch with the contact person for this course.

Target Audience:

The course is primarily aimed at HPC support staff who would like to help users run and optimise machine learning workloads. It is also suitable for participants with HPC experience interested in implementing and profiling their own training workflows.

Language:

The course is given in English.

Duration:

2 days

Dates:

17.-18. November 2026, 09:00 - 12:00 and 13:00-17:00 each day

Venue:

Jülich Supercomputing Centre, Building 16.4, Rotunda

Number of Participants:

Minimum 5, maximum 30 participants

Instructors:

Stefan Kesselheim, Sabrina Benassou, Javad Kasravi, Eske Ewert