For most of its history, CINECA could design supercomputers around a relatively well-understood set of priorities. Scientific simulation demanded enormous compute capability, fast interconnects and parallel storage capable of sustaining large sequential reads and writes.
The architecture was complicated, certainly, but the workload model was familiar. AI has disrupted that model remarkably quickly. Daniele Cesarini, Head of AI/HPC Architecture at CINECA, has watched that change happen directly on Leonardo, the EuroHPC pre-exascale system that remains one of Europe’s largest supercomputers.
Cesarini broke this and more down for us on the latest episode of the Shared Everything podcast.
A Very Different Kind of I/O Problem
Just a few years ago, he says, roughly 95 to 98 percent of Leonardo’s workload was traditional HPC. Today, more than half is AI. That shift changes more than the balance between CPUs and GPUs. It changes what the infrastructure underneath the compute has to do.
Traditional HPC storage was built around applications that tend to read and write large amounts of data through high-performance parallel file systems. AI introduces a much broader and less predictable collection of data-access patterns. CINECA still needs high-performance file access for simulation Cesarini says, but those same environments now have to support AI training, containerized applications, cloud infrastructure and increasingly inference.
He says that data may need to be presented through a parallel file interface for one workload, object for another and NFS somewhere else but for Cesarini, that meant the old assumption that a high-performance parallel file system could serve as the center of the storage architecture no longer held.
The issue is not simply storage performance, he explains. It’s the ability to maintain that performance while supporting very different interfaces and workloads against a common data platform. CINECA already operates this way across its infrastructure. Its data platform can be mounted into OpenStack cloud environments while also serving Leonardo and AI-focused infrastructure.
Cesarini admits containerized applications require their own access mechanisms, while AI services introduce still more requirements, including capabilities around vector databases and KV caching and that creates a very different architectural problem from simply feeding a simulation as quickly as possible.
From Storage System to Data Platform
This is where CINECA’s relationship with VAST Data becomes important. The organization is already operating VAST across multiple environments, including roughly 50 PB associated with its Bologna cloud infrastructure and another 50 PB at a second datacenter.
Its upcoming Italia AI Factory will add roughly another 100 PB-scale data platform. Capacity, however, is only one part of the architectural story. What matters to Cesarini is being able to place different access methods and services around the same underlying data. CINECA can retain the high-performance file access its HPC applications require while exposing object interfaces, NFS and CSI-based access for cloud and container environments.
That becomes increasingly important as the distinction between those environments starts disappearing. An HPC center can no longer assume that a workload begins and ends inside the supercomputer. Data can move through cloud services, preprocessing, training and inference while being consumed by applications with very different expectations about how that data should be presented.
Cesarini describes the resulting environment as highly heterogeneous, and that heterogeneity is now an architectural requirement rather than something to be engineered away.
Inference Changes the Architecture Again
Training is only part of the change, he also says. CINECA is beginning to provide inference endpoints and is considering what Cesarini describes as token-factory services running on its infrastructure. That pushes the data layer into another territory traditionally far removed from HPC storage. KV caching, for example, becomes important when trying to increase token-generation rates during inference, while vector database capabilities become relevant as AI applications retrieve and operate on data rather than simply consuming datasets during a training run. These aren’t capabilities that historically appeared on the requirements list for a supercomputing parallel file system, yet it’s not like CINECA can simply replace its HPC architecture with an AI architecture. Its traditional scientific workloads remain part of its mission, so the harder requirement is to support both simultaneously without creating separate data silos for every new workload class.
Feeding More Than 8,000 GPUs
Italia is one of the European AI Factories being developed to expand sovereign AI capacity across the region.
CINECA expects the system to contain more than 8,000 GPUs connected through NVIDIA InfiniBand and Ethernet, with VAST providing the underlying data platform. At that scale, adding GPUs is only useful if the rest of the system can keep them productive, making the data architecture a direct part of the compute-efficiency problem.
Italia therefore represents the next stage of the transition CINECA has already experienced on Leonardo. Instead of adding AI to infrastructure originally dominated by HPC, the organization can design around mixed AI, HPC and cloud requirements from the beginning, and those requirements extend beyond throughput.
As CINECA works with private companies alongside universities and public research institutions, Cesarini says security and multi-tenancy have become substantially more important. Traditional supercomputing environments could optimize overwhelmingly for performance.
Shared AI infrastructure has to preserve that performance while isolating tenants, supporting different interfaces and allowing data to move through much more complicated application pipelines. The architecture starts looking less like a conventional supercomputer with storage attached and more like a shared computing environment in which the data layer connects multiple forms of compute.
The longer-term problem is even more interesting because Cesarini does not expect HPC and AI to remain cleanly separated workload categories. AI is already being inserted into scientific workflows, creating pipelines in which simulation and AI operate together during execution. CINECA is also beginning to think about training across multiple datacenters. It currently operates three, raising the possibility of distributing workloads across sites while coordinating them as a single computational job, which compounds the data problem.
The same data may eventually need to support HPC, AI and cloud stages of a workflow across multiple locations, with different access protocols, security requirements and tenancy models at each stage. All of it still has to deliver the predictable performance expected from a supercomputing environment. For Cesarini, maintaining that performance stability while combining these requirements is one of the major architectural challenges ahead.
The interesting lesson from CINECA is therefore not simply that AI requires faster storage. It requires a different conception of what the data layer of a supercomputer is supposed to be. Leonardo showed how quickly the workload mix can change, while Italia is being built around what comes next: thousands of GPUs, HPC and AI operating side by side, inference joining training, cloud services intersecting with supercomputing and potentially workloads spanning multiple datacenters. Compute remains the visible measure of these machines, but increasingly, the harder architectural question is what sits underneath it and whether the data platform can keep every part of that increasingly heterogeneous system fed, connected and running at full speed.
