AI Is Bringing Added Complexity for HPC Sites. How Are They Handling It?

Key Takeaways

• CINECA added 30 MW of power across two new data centers
• Staff grew from 70 to roughly 200 to handle AI complexity
• Support now spans Slurm, Kubernetes, OpenStack, and multiple storage types
• Users demand thousands of AI libraries and containerized pipelines

Pulse Analysis

The arrival of artificial intelligence in high‑performance computing has upended the traditional HPC playbook. Where sites once relied on a narrow set of numerical libraries, a single workload manager and parallel file systems, today’s researchers expect deep learning frameworks, vector databases and streaming platforms. This diversification forces administrators to juggle PyTorch, TensorFlow, Spark, Kafka and more, while ensuring compatibility with legacy MPI codes. The resulting software sprawl demands robust orchestration, observability and security layers that were previously unnecessary.

CINECA’s response illustrates how a leading European consortium is scaling both hardware and human capital. After installing Leonardo—now ranked in the TOP500—the organization launched two additional 30‑megawatt data centers, planning three more AI systems, two HPC machines and a regional cloud. Employee headcount tripled to about 200, with new roles specializing in GPU storage, Kubernetes, OpenStack and network fabrics. By adopting VAST’s data platform, CINECA can offer file, block and S3‑compatible storage, reducing bottlenecks for data‑intensive AI pipelines.

The broader lesson for HPC operators is clear: the era of a monolithic stack is over. Sites must invest in multi‑cloud strategies, container orchestration and a library‑as‑a‑service model to stay competitive. As AI workloads proliferate, the ability to provision bespoke environments quickly will become a differentiator, influencing funding decisions and research collaborations worldwide.