This week, Hewlett Packard Enterprise announced that it had landed a contract from the EuroHPC Joint Undertaking (EuroHPC JU) consortium to build the LUMI (Large Unified Modern Infrastructure) supercomputer in Kajaani, Finland. The machine will be based on AMD's next-generation EPYC processors as well as Instinct compute accelerators, with peak performance of more than 0.55 ExaFLOPS (550 PetaFLOPS, 550K TFLOPS).
The LUMI Supercomputer
Just like its more powerful exascale-class supercomputers from the U.S., the LUMI supercomputer will use HPE's Cray EX architecture powered by AMD's upcoming EPYC 'Milan' CPUs as well as next-generation Instinct GPUs. The system is expected to have a maximum performance of more than 375 PetaFLOPS and a peak performance of more than 550 PetaFLOPS. Today's world's most powerful Fugaku supercomputer features maximum performance of 415.5 PetaFLOPS as well as peak performance of 513.8 PetaFLOPS.
Based on EuroHPC's projections, LUMI will be among the top 5 supercomputers in the world when it is fully deployed by late 2021.
LUMI takes over 150 m2 of space, which is approximately the size of a tennis court. The supercomputer will consist of multiple partitions designed for specific tasks and interconnected using HPE's Slingshot technology to combine different resources if needed.
- LUMI-G is the key partition hosting AMD Instinct GPUs that will deliver a peak performance of over 550 PetaFLOPS.
- LUMI-C is a supplementation partition that hosts ~200,000 of AMD EPYC cores.
- LUMI-D is an interactive partition with 32 TB of memory and graphics processors for data analytics and visualization.
- LUMI-F serves as an accelerated all-flash-array storage partition based on the Cray ClusterStore E1000 that that can store 7 PB of data while providing up to 2 TB/s of aggregated bandwidth.
- LUMI-O 90 PB encrypted object storage partition.
- LUMI-P 80 PB Lustre parallel file system storage partition.
- LUMI-K contained cloud service partition.
- LUMI-Q emerging technology partition.
LUMI will be used for scientific research in artificial intelligence, climate change, medicine, weather forecasting, pharmaceutical discovery, and others.
The system will consume approximately 8.5 MW when fully built and powered by a hydropower station that can output up to 200 MW. To that end, the system can be considered carbon neutral. It is necessary to note that the system will be particularly energy-efficient at 44.1 GigaFLOPS per Watt, which is two times more efficient than MN-3, the current Green500 list leader.
The LUMI supercomputer will be hosted at CSC – IT Center for Science in Finland and will be jointly used by public and private researchers from ten countries, including Finland, Belgium, Czech Republic, Denmark, Estonia, Iceland, Norway, Poland, Sweden, and Switzerland. The total budget of the product, which will be co-funded by the EuroHPC JU and the LUMI Consortium, will total €144.5 million (over $165 million). The total cost of ownership of the system from 2020 to 2026 will be €200 million ($237.1 million), which includes the cost of hardware.