Berzelius is an NVIDIA SuperPOD consisting of 94 DGX-A100 nodes, sporting a total of 752 NVIDIA A100 GPUs.
The SuperPOD uses the SLURM resource manager and job scheduler. The original 60 DGX-A100 nodes have 8x NVIDIA A100 GPUs (40GB), 128 CPU cores (2x AMD Epyc 7742), 1 TB of RAM and 15 TB of NVMe SSD local disk. The 34 newer DGX-A100 nodes have 8x NVIDIA A100 GPUs (80GB), 128 CPU cores (2x AMD Epyc 7742), 2 TB of RAM and 30 TB of NVMe SSD local disk. High performance central storage is available using 4x AI400X and 2x AI400X2 from DDN, serving 1.5 PB of storage space to all nodes of the cluster. All DGX-A100 GPUs have dedicated Mellanox HDR InfiniBand HBAs, that is, there are 8 Mellanox HDR HBAs per DGX-A100 node, connected in a full bisection bandwidth, fat-tree topology.
More information about Berzelius Compute @ NSC
Open Round
At this moment, you can apply for access to this resource in the following round:
LiU Berzelius 2024
Apply to this round for a project on the Berzelius SuperPOD at NSC. Berzelius is a non SNIC system donated to NSC by Knut and Alice Wallenberg foundation (KAW). Monthly evaluation of proposals during the year.
To apply, you must be a scientist in Swedish academia, at the level of PhD student or higher.
Applications are normally evaluated during the last week each month
See further information.