Sweden’s National Supercomputer Center today announced the launch of Berzelius, a supercomputer based on Nvidia’s DGX SuperPOD architecture and capable of delivering 300 petaflops of AI performance, using Nvidia’s half-precision tensor cores. The new system, built by Atos, is named for Jacob Berzelius, the Swedish chemist best known for developing chemical notation – think H2O as shorthand for water – and discovering several elements including, appropriately, silicon.
Berzelius is expected to become Sweden’s fastest supercomputer as measured on the next Top500 list (June), more than doubling the performance of the two prior top Swedish systems. Extrapolating from the performance of Nvidia’s Selene system, which is also based on the SuperPOD architecture, a reasonable estimate is that Berzelius will deliver more than 6 petaflops Linpack – that’s based on Selene’s Top500 scores. Presumably more details will be available in June when the next Top500 results are issued.
Funded by a €29.5 million donation from the Knut and Alice Wallenberg Foundation, Berzelius is part of Sweden’s NCS, which is located at Linköping University and located roughly 70 miles from where Berzelius was born. “Many cities in Sweden have a square or street that bears Berzelius’s name, but the average person probably doesn’t know much about him,” said Niclas Andersson, technical director, NCS, in a Nvidia blog announcing the launch.
The Berzelius system consists of 60 Nvidia DGX A100 systems, linked on a 200 Gbps Nvidia Mellanox InfiniBand HDR network. The same network links the processors to 1.5 petabytes of flash memory on four storage servers from DataDirect Networks (DDN AI400X with NVMe).
“Buying more storage might be the first upgrade we will need because 60 DGX A100 systems can suck up a lot of data fast,” said Andersson in the blog. To get started quickly, NSC asked Nvidia and Atos, who managed the build and integration, to configure the system’s software. The stack will include the Atos Codex AI Suite as well as access to NGC, Nvidia hub for GPU-optimized software for AI and HPC.
NSC reports the researchers who will primarily work with the new supercomputer are associated with the programs funded by the Wallenberg Foundation, such as the Wallenberg AI Autonomous Systems and Software Program (Wasp). Early projects will target advancements in areas such as “wireless communications, cybersecurity, large-scale IoT, and efficient programming” according to NSC.
“Research in machine learning requires enormous quantities of data that must be stored, transported and processed during the training phase. Berzelius is a resource of a completely new order of magnitude in Sweden for this purpose, and it will make it possible for Swedish researchers to compete among the global vanguard in AI,” said Anders Ynnerman, professor of scientific visualization at Linköping University and program director for Wasp.
“We are really delighted to have been working with Linköping University on the delivery and installation of this new high-performance supercomputer. With Berzelius, researchers will now have powerful computing power that is able to harnesses the power of deep learning and analytics, in order to speed-up data processing times, and provide researchers with insights faster, thereby helping Sweden to address some of the key challenges in AI and machine learning today,” said Pierre Barnabé, senior executive VP and head of the Big Data and Cybersecurity Division, Atos.
“We are extremely happy for research in Sweden that the Wallenberg Foundations have been able to contribute to the acquisition of world-class computer infrastructure in a location that supplements and reinforces the major research initiatives we have made in recent years in such fields as AI, mathematics and the data-driven life sciences,” said Marcus Wallenberg, vice-chair of the Wallenberg Foundation.
Linköping University release, https://liu.se/en/news-item/sveriges-snabbaste-superdator-for-ai-ar-invigd