Our supercomputers

Compute partitions:

LUMI’s storage system consists of three components. First, there is a 7-petabyte partition of ultra-fast flash storage, combined with a more traditional 80-petabyte capacity storage, based on the Lustre parallel filesystem, as well as a data management service, based on Ceph and being 30 petabytes in volume. In total, LUMI has a storage of 117 petabytes and a maximum I/O bandwidth of 2 terabytes per second.

LUMI takes over 150m2 of space, which is about the size of a tennis court. The weight of the system is nearly 150 000 kilograms (150 metric tons)

GPU-partition (Booster) delivering 240 petaflops

x86 CPU-partition (Data-Centric) delivering 9 petaflops and featuring DDR5 Memory and local NVMe.

13824 "Da Vinci" GPUs (based on NVIDIA Ampere architecture) delivering up to 10 exaflops of FP16 Tensor Flow AI performance

All the applications suit ideally MareNostrum5 thanks to its heterogeneous configuration, with a special focus on medical applications, drug discovery as well as digital twins (earth and human body), energy, etc. Its large general-purpose partition provides an environment well suited for most current applications that solve scientific/industrial problems. In addition, the accelerated partition provides an excellent environment for large scale simulations, AI and deep learning.

MareNostrum 5 is located in BSC’s new facilities, next to the Chapel which is hosting previous systems. The datacenter has a total power capacity of 20MW, and cooling capacity of 17MW, with a PUE below 1,08.

CPU partition: 960 nodes with 2CPUs and 256GB memory/node (20% 1TB/node), 1x HDR100 & GPU partition: 60 nodes with 2CPUs and 512GB memory, 2x HDR100, 4x Nvidia A100/node

240x Nvidia A100 with 40 GB HBM2 (+4 on GPU login nodes), 6912 FP32 CUDA cores and 432 Tensor cores per GPU

6x 100 Gbit/s bandwidth for data transfers to other national and international computing centres, data processing throughput of more than 400GB/s with high-performance storage and 200GB/s with large-capacity storage

20 PetaBytes main storage with all-flash scratch tier over 600GB/s, Tape archival capabilities

The supercomputer consists of 6 main parts:  

• a universal part for standard numerical simulations, which will consist of approximately 720 computer servers with a theoretical peak performance of 3.8 PFlop/s,
• an accelerated part with 70 servers and each of them being equipped with 8 GPU accelerators providing a performance of 11 PFlop/s for standard HPC simulations and up to 150 PFlop/s for artificial intelligence computations, 
• a part designated for large dataset processing that will provide a shared memory of as high as 24 TB, and a performance of 74 TFlop/s,
• 36 servers with a performance of 131 TFlop/s will be dedicated for providing cloud services,
• a high-speed network to connect all parts as well as individual servers at a speed of up to 200 Gb/s,
• data storages that will provide space for more than 1 PB of user data and will also include high-speed data storage with a speed of 1 TB/s for simulations as well as computations in the fields of advanced data analysis and artificial intelligence.

A64FX (ARM partition), AMD EPYC (x86 partitions)

430 TB High-speed NVMe partition, 10.6 PB high-speed based Parallel File System partition.

Deucalion will be installed at the Portuguese Foundation for Science and Technology (FCT) Minho Advanced Computing Centre (MACC), in close collaboration with the municipality of Guimarães, in the North of Portugal, as part of a fully sustainable computing infrastructure aiming at promoting new advancements in the digital and green transitions.