A hybrid Slurm-Kubernetes platform to integrate AI workloads in HPC infrastructure

The exponential growth of artificial intelligence (AI) is reshaping the computational landscape, placing unprecedented demands on supercomputing centers and SMEs alike. While traditional High-Performance Computing (HPC) environments rely on workload managers like Slurm to orchestrate large-scale, deterministic simulations, AI workloads introduce non-deterministic models, heterogeneous software stacks, and real-time inference demands that challenge conventional HPC architectures.At the same time, Kubernetes (K8s) has emerged as the de facto standard for scalable AI deployment and microservices orchestration, yet its performance isolation, GPU scheduling efficiency, and integration into HPC resource management remain limited. Bridging these paradigms is essential for the next generation of hybrid computing infrastructures. This talk presents the design and implementation of a hybrid Slurm–Kubernetes platform developed by Do IT Now to integrate AI workloads seamlessly within HPC infrastructures.The proposed architecture enables dynamic compute node allocation between Slurm and Kubernetes, managed under a unified and performance-oriented infrastructure. Unlike isolated deployments, the hybrid model allows HPC users to run traditional batch simulations and containerized AI workflows on the same hardware, optimizing utilization and cost efficiency. The current solution supports Slurm workload manager on-premises and cloud bursting modes, and preserves traditional HPC-grade multi-tenancy and accounting capabilities, which represents a key requirement for research centers and AI factories operating under strict security and compliance frameworks. Results from pilot deployments demonstrate significant gains in cluster efficiency and GPU utilization, reducing idle times while maintaining the deterministic performance expected in HPC environments.From an operational perspective, the unified interface simplifies user onboarding and reduces the steep learning curve associated with MLOps workflows, enabling data scientists and researchers to focus on model development rather than infrastructure management. In conclusion, this hybrid Slurm–Kubernetes platform offers a pragmatic path toward converging HPC and AI operations, promoting scalability, reproducibility, and cost-efficiency across both domains.ABOUT THE AUTHORSMarco De La Pierre is a Computational Scientist at Do IT Now, a global HPC services and consulting company. Building on a background in computational materials science and on 15 years of experience in High-Performance Computing, he enjoys investigating new software technologies to design and deploy HPC solutions that facilitate user experience and productivity. Active across various aspects of applications and user support, his interests include workflow automation, reproducibility and portability, scientific software stack management, and user training.Jordi Blasco Pallares is a computational physicist from the University of Barcelona with over two decades of experience in HPC. His career has spanned multiple national-level supercomputing centers and international research institutions, where he has architected, optimized, and managed large-scale computing systems supporting scientific and industrial innovation. As Co-Founder and Director of Do IT Now - New Zealand (formerly HPCNow!) Jordi leads strategic consulting and solution design across the APAC region and the West Coast of the USA. His work focuses on the development and deployment of advanced HPC and AI infrastructures. He has played a key role in enabling research and engineering teams to achieve peak computational efficiency through performance tuning, benchmarking, and automation. -----For more information about the eResearch NZ / eRangahau Aotearoa conference, visit: https://eresearchnz.co.nz/