Beyond Supercomputer

Imagine standing at an airport, looking out across the runway as one plane after another makes its approach and comes in to land. For a long time, scientific computing on supercomputers worked in much the same way. If you wanted to run a weather model or study the effects of new drugs, you were allocated a time slot. And your work went ahead provided the machine had enough capacity.

Today, things are different. A modern supercomputer such as Alps, inaugurated at the Swiss National Supercomputing Centre (CSCS) in 2024, is a long way from being a single runway. It is more like an entire airport, with several runways and many terminals. This greater complexity is necessary, because researchers are running more large-scale data analyses, simulations and AI models. At the same time, the data, software tools and workflows involved vary from one discipline to the next. A supercomputing infrastructure must therefore meet a host of different requirements. Fast hardware and high computing power alone are no longer enough.

From computer to infrastructure

Alps marks CSCS's completed transition away from monolithic computers to a cloud-native research infrastructure. The centre now provides researchers at ETH Zurich and across Switzerland with additional tools, platforms and services tailored to the needs of their disciplines. Whatever their requirements, they can all use the same infrastructure. But, to return to the airport analogy, they do not have to use the same runway - or even the same terminal.

CSCS calls this concept "versatile architecture". "Researchers used to adapt to the supercomputer; today we adapt the supercomputer to researchers and their needs," says Maria Grazia Giuffreda, associate director at CSCS, responsible for the user programme and for collaboration with researchers.

"We want Alps to be as flexible to use as possible," adds Joost VandeVondele, deputy director for science at CSCS, who is responsible for the technical development of the research infrastructure. "The technology behind high-performance computing should be smooth and streamlined, empowering researchers to focus on their research rather than the underlying infrastructure."

With Alps, CSCS built an integrated system in which hardware, software and specialist services mesh across several layers, each building on the one below. The computing system forms the foundation, with software layers and cloud services on top. Together, they make it easier for researchers to access computing power, data and scientific applications.

The vClusters are a case in point. This additional software layer divides the Alps hardware into flexible areas, allowing ETH researchers and other users such as MeteoSwiss to work with Alps in their own way. CSCS's cloud services, web portals, data platforms and AI applications all sit on top of this software layer. Since July 2026, for example, researchers have been able to use the Apertus 1.5 large language model through a new service, with Alps automatically supplying the required computing power.

Driven by AI

CSCS originally developed its supercomputers primarily to perform large-scale scientific simulations. Today the driving force behind supercomputing development is artificial intelligence. From the outset, Alps was designed so that simulations, data analyses and AI applications could all run on the same system.

"Thanks to its versatile architecture, Alps brings two worlds together: the precise calculations of classical simulations and the data-intensive processes of modern AI," says Torsten Hoefler, an ETH professor of computer science specialising in high-performance computing and chief architect for AI and machine learning at CSCS. "What sets us apart is that we don't just use the hardware - we also develop algorithms, software and infrastructure that are designed to work together."

The Alps hardware is not built at CSCS. But thanks to close collaboration with the manufacturers, CSCS was the launch customer of the Grace Hopper Superchips, which form a key component of the supercomputer. Those chips are highly sought-after for today's AI applications because they process data very quickly.

"The GPUs are a key component of these chips. CSCS embraced GPU technology early on to accelerate computing operations," says Giuffreda. "This meant we could contribute our experience, and we were among the first to use these chips." For Hoefler, the timing was decisive: "Thanks to this early investment, CSCS and the ETH Domain were ready when the big AI wave broke."

This makes Alps a cornerstone of Switzerland's infrastructure for AI research. The Swiss AI Initiative, launched by ETH Zurich and EPFL in 2023, uses Alps to develop foundation models, for instance. These AI models are trained on very large volumes of data and then put to work in fields such as climate, weather and health research.

The training takes place on Alps; after that, researchers can use the models on other computers and adapt them as required. In future, CSCS intends to further expand its AI services beyond model training. Already today, CSCS offers a service that provides researchers easy access to open-source AI models for their own projects and applications without having to install the models on their computer.

Intelligence beneath the surface

A researcher whose expertise contributes to this user service is Ana Klimovic. The ETH Zurich professor of computer science investigates the system software that operates in a layer between applications and hardware. Together with CSCS, she is currently working to expand the service so that, in future, researchers will be able to access not only the computing resources of Alps, but also those of other supercomputing centres abroad. This would provide substantially greater computing capacity, particularly for demanding AI applications.

Klimovic's group also explores how to make more efficient use of computing resources in supercomputers. The group developed Sailor, an automated AI training platform. It helps researchers select the most suitable computing resources for their AI jobs and make the most efficient use of the available hardware.

"The more effectively we design software to make use of the hardware resources and available power in Alps, the more quickly each job in the cluster completes and hence the more jobs we get to run on Alps to explore all kinds of research goals," says Klimovic and adds, "Alps allows my research group to develop, test and use new software on state-of-the-art AI hardware to improve performance and energy-efficiency." In this respect, Klimovic's software acts like the air traffic controllers who assign runways to incoming flights.

"It's not just the hardware infrastructure that makes Alps so special. The supercomputer enables CSCS and researchers to create new knowledge and new technologies," says Giuffreda. "That's vital to Switzerland's digital sovereignty."