The Flemish Supercomputer Center (VSC) is preparing for the next Tier-1 supercomputer. It will be located at the VUB, but will be available to all scientists.
The machine (in practice several connected machines) will be operational around October-November 2025 from the new data hub of the VUB, the Green Energy Park site in Zellik. The investment amounts to 12 million euros.
It is not yet known who will build the Tier-1 supercomputer and with what specifications. “It has now been decided that we will house the machine, now the European tender will follow,” Ward Poelmans, department head of Scientific Data and Compute at the VUB, tells Data News. That tender will be completed at the end of next year or early 2025.
Despite the hype around AI and language models, the new machine remains very general. This is because, partly due to the enormous cost, it will be used by all Flemish universities.
Today they each have a Tier-2 supercomputer, but for Tier-1 machines there is collaboration. For example, Ghent University has currently had Hortense since 2021, the Tier-1 supercomputer that was purchased a few years ago.
To give an idea of the capacity at the time: Hortense has peak performance of up to 3.3 petaflops, which equates to 3,300,000,000,000,000 operations per second. For this it uses 44,000 CPU cores, 100 terabytes of RAM and 3 petabytes of storage capacity.
‘The supercomputer in Ghent is quite classic. But when we ask researchers what they need, it is actually more of the same,’ Poelmans explains. ‘The majority is still based on classic CPUs. There is also a reasonable component of GPUs and that will probably increase. But it is also a matter of consideration because the new generation of GPUs are particularly expensive,” says Poelmans.
At the same time, Poelmans nuances that although a supercomputer can also be used to train language models or other AI applications (which often rely mainly on the GPU), Flemish scientific research goes much broader. ‘If we look at the use of those GPUs today, it is not mainly for machine learning, but also for quantum chemistry, for example, because that is code that runs much faster on a GPU.’
Training a language model (large language model or llm) such as that of OpenAI or Meta is also too big to repeat in Flanders. ‘That is gigantic in scale and doing so would cost millions of euros. What does happen in Flanders is retraining for specific applications.’
In general, Poelmans expects that the new generation of Tier-1 supercomputer will be used more often for machine learning and will rely relatively more on GPU computing power, but certainly not mainly. ‘The Flemish universities are a very diverse group and you want to serve everyone. Therefore, it will be a compromise for every discipline.”