Berzelius is Sweden’s fastest supercomputer for AI and machine learning

There is no doubt that artificial intelligence (AI) is something the technology giants are investing heavily in. AI calculations are performed through the Tensor cores known to SweClocker members in Nvidia’s “Turing” and “Ampere” architectures, which are used, for example, for scaling up with DLSS. Defeating human world champions in computer games and optimizing the use of energy in laptops are other AI-related activities.

Swedish researchers in the field of artificial intelligence and machine learning are now getting a really upgraded toolbox, in connection with the inauguration of Sweden’s most powerful supercomputer for AI. The supercomputer takes place in the server halls at the National Supercomputer Center (NSC) at Linköping University, which opens the door to a theoretical computing power of 300 petaflops at mixed-precision.

In true supercomputer spirit, the system has a name: Berzelius, after the prominent scientist of the Östgötaslätten Jacob Berzelius – “the father of Swedish chemistry”. Under the shell of the modern version of Berzelius is water-cooled hardware in the form of 60 Nvidia-signed boxes of the type DGX Superpod, with the graphics circuits A100 in the center. Each box is populated by 128 processor cores AMD “Rome” and eight “Ampere” -based top circuits equipped with 40 GB of memory each, which with 480 circuits will be just over 19 TB. At full load, said water cooling can dissipate heat of about 300 kW.

Each DGX box is equipped with ten fiber connections – eight at 200 Gbps and the remaining at 100 Gbps – and communication between the subsystems takes place via Mellanox hardware. Part of Berzelius are also four NVME-equipped storage servers optimized for use in AI contexts, which after redundancy have a capacity of 1 petabyte (PB). Total read and write performance is at a breathtaking 192 and 136 GB / s respectively. On the software front, the company Atos is responsible for a solution that “supports researchers in using the system efficiently”.

Like many supercomputers, it is purely structurally about a series of server cabinets, where a side panel with Berzelius’ name marks what it is about. Appropriately, the names and logos are accompanied by the “AI” binary ASCII code. The porky hardware comes with an equally porky price tag of SEK 150 million, where the double amount was donated from the Knut and Alice Wallenberg Foundation to also allow expansion in a few years.

Research in machine learning requires enormous amounts of data that must be stored, transported and processed during the training phase. Berzelius is a resource of a new order of magnitude in Sweden for this purpose that will enable Swedish researchers to compete on the absolute world front in AI

Anders Ynnerman, who is a professor of visualization, believes that the very heavy training phase in the AI ​​context can now be handled and take Swedish research to the forefront of the subject area. In conversations with NSC’s technical manager, Niclas Andersson, it is explained that the future will show exactly how researchers will use the tool, but that the use will probably be more dynamic than with the other supercomputers in NSC’s arsenal.

Andersson also says that the future includes an update of the Top 500 list for supercomputers, which KTH also has in mind with the recently unveiled supercomputer Dardel. The list will be updated in June and Niclas Andersson predicts that it is possible that Berzelius will end up somewhere in the 60-70 range.

Source: Linköping University