Integrating Artificial Intelligence and High-Performance Computing

Researchers in fields such as linguistics, medicine, economics, and environmental studies increasingly rely on AI methods; and as models and datasets grow in scale, access to High-Performance Computing (HPC) and optimized AI workflows becomes a necessity. Many best practices in machine learning and AI are scattered across tools and papers, HPC systems differ in hardware, schedulers, and file systems, which complicates model choice, resource allocation, and reproducible workflows. The goal of this poster is to provide a simple, structured overview that helps researchers plan and run AI projects on HPC systems. This poster is organized into the following subsections, which encorporate the essential aspects of AI on HPC: Hardware stack. The main node components—compute, memory, storage, network—and their impact on performance. Software stack. Core workflow — SLURM scheduling, environments, MPI configuration, logging and checkpointing. Data & file systems: Efficient data handling using parallel file systems and object stores, with basic guidelines for staging or caching to improve I/O performance.. Frameworks: Frameworks to facilitate training and inference, e.g., PyTorch/TensorFlow/JAX for training, ZeRO/FSDP for parallel computing and sharding, and inference servers like TorchServe, TF Serving for end-to-end AI on clusters Performance: Tools and guides to monitor training and inference efficiency on HPC using profiling tools to track GPU usage, memory, and scaling behavior for understanding resource utilization and runtimes. Alongside the technical content, the poster also shows how HPC.NRW enables researchers to (i) identify local HPC centers and avail resources, (ii) access training on HPC and AI topics through courses and HPC-Wiki, and (iii) request support and consultations for their AI projects. The poster is designed as a quick reference. Conference participants—and later readers—can rapidly see the major steps of the AI project lifecycle on HPC: select models, prepare data, choose resources, run, measure, and iterate. Finally, we invite researchers to engage with HPC.NRW as partners for scaling AI projects on real clusters.