Accelerator-based resource dedicated to research using AI techni
The Alvis cluster is a national NAISS resource dedicated to Artificial Intelligence and
Machine Learning research.
Note: Significant generation of training data is expected to be done elsewhere.
The system is built around Graphical Processing Units
(GPUs) accelerator cards. The first phase of the resource has 160 NVIDIA T4, 44
V100, and 4 A100 GPUs. The second phase is based on 340 NVIDIA A40 and 336
A100 GPUs.
LUMI is a general computational resource hosted by CSC in Finland.
LUMI, Large Unified Modern Infrastructure, is an HPE Cray EX supercomputer consisting of several partitions targeted for different use cases. The largest partition of the system is the LUMI-G partition consisting of GPU-accelerated nodes using AMD Instinct GPUs. In addition to this, there is a smaller CPU-only partition LUMI-C that features AMD Epyc CPUs and an auxiliary partition for data analytics with large memory nodes and some GPUs for data visualization.
The LUMI consortium countries are Finland, Belgium, the Czech Republic, Denmark, Estonia, Iceland, Norway, Poland, Sweden, and Switzerland. The acquisition and operation of the EuroHPC Supercomputer are jointly funded by the EuroHPC Joint Undertaking and the LUMI consortium. The Swedish Research Council has contributed approx. 3.5% of the funding. A corresponding share of the system is reserved for Swedish research, but researchers are encouraged to apply for resources from the JU part of LUMI and other EuroHPC resources.
LUMI is a general computational resource hosted by CSC in Finland.
LUMI, Large Unified Modern Infrastructure, is an HPE Cray EX supercomputer consisting of several partitions targeted for different use cases. The largest partition of the system is the LUMI-G partition consisting of GPU-accelerated nodes using AMD Instinct GPUs. In addition to this, there is a smaller CPU-only partition LUMI-C that features AMD Epyc CPUs and an auxiliary partition for data analytics with large memory nodes and some GPUs for data visualization.
The LUMI consortium countries are Finland, Belgium, the Czech Republic, Denmark, Estonia, Iceland, Norway, Poland, Sweden, and Switzerland. The acquisition and operation of the EuroHPC Supercomputer are jointly funded by the EuroHPC Joint Undertaking and the LUMI consortium. The Swedish Research Council has contributed approx. 3.5% of the funding. A corresponding share of the system is reserved for Swedish research, but researchers are encouraged to apply for resources from the JU part of LUMI and other EuroHPC resources.
General purpose non-accelerated computational resource
Arrhenius CPU contains two CPU node types, 192 thin nodes, and 20 fat nodes. Each node have 256 cores split between two CPU:s. The thin nodes have 768 GB RAM and the fat nodes have 3072 GB RAM. There is a 1.92 TB local SSD on each node. It is accompanied with the Arrhenius Disk and Arrhenius Flash storage systems.
Arrhenius CPU is primarily intended for work that cannot be moved to the much more powerful accelerated computational resource, the Arrhenius GPU. Projects with such needs will be prioritized on Arrhenius CPU.
Arrhenius GPU contains 382 GPU nodes. Each GPU node have 4x Nvidia GH200 GPU:s with 96GB HBM and 128 GB LPDDR Modules. There is a 1.92 TB local SSD on each node. Arrhenius GPU is accompanied with the Arrhenius Disk and the Arrhenius Flash storage systems.
Arrhenius GPU provide the main computational power in the Arrhenius HPC system.
Arrhenius SENS Compute is a resource suitable for projects requiring an increased security level, notably sensitive personal data such as whole human genomes associated with living individuals.
The resource includes CPU and GPU hardware. To simplify administration, GPU usage will be accounted in core-hours at a rate of 50 core-hours per GPU-hour for an Nvidia L40S and 200 core-hours per GPU-hour for an Nvidia H100.
It contains:
Small CPU nodes, each with 2x 128-core AMD EPYC 9745 at 2.4 GHz and 768 GB RAM.
17 Large CPU nodes, each with 2x 128-core AMD EPYC 9745 at 2.4 GHz and 3072 GB RAM.
20 Small GPU nodes, each with 4x Nvidia L40S 48GB GPU's and 2x 32-core AMD EPYC 9335 at 3.0 GHz and 384 GB RAM.
1 Large GPU node, with 4x Nvidia H100 80GB and 2x 32-core AMD EPYC 9335 at 3.0 GHz and 768 GB RAM
Tetralith is a general computational resource hosted by NSC at Linköping University.
Tetralith servers have two Intel Xeon Gold 6130 processors, providing 32 cores per server. 1844 of the servers are equipped with 96 GiB of primary memory and 64 servers with 384 GiB. All servers are interconnected with a 100 Gbit/s Intel Omni- Path network which is also used to connect the existing storage. Each server has a local SSD disk for ephemeral storage (approx. 200GiB per thin node, 900GiB per fat node). An IBM Spectrum Scale system comprises the centre storage. 170 of the Tetralith nodes are equipped with one NVIDIA Tesla T4 GPU each as well as a high- performance NVMe SSD scratch disk of 2TB.
Dardel is a Cray EX system from Hewlett Packard Enterprise, based on AMD EPYC processors with an accompanying Lustre storage system.
The nodes are interconnected using Slingshot HPC Ethernet.
Dardel-GH is the accelerated partition based on nVIDIA GraceHopper GPUs. It has an accompanying Lustre storage system. The nodes are interconnected using Slingshot HPC Ethernet.
A new high-performance GPU resource is now accessible on the Dardel supercomputer at PDC, providing 20,000 GPU hours per month. Eight additional nodes have been purchased, each equipped with four NVIDIA Grace Hopper (H100) GPUs and four NVIDIA Grace ARM CPUs—totalling 32 Grace Hopper Superchips. This state-of-the-art platform combines the performance of the Hopper GPU with the energy-efficient Grace ARM CPU, designed for demanding AI and high-performance computing (HPC) workloads.
The Grace Hopper resource is exclusively available to projects under the Medium compute category, with priority given to users running codes not currently supported on Dardel's AMD GPUs. If there is a need to run unique or specialized software, this new system offers a significant opportunity for improved performance.
Important:
For users with an existing Medium project on Dardel with a need to use dardel-GH resource, submit a continuation of your existing Medium project (you can submit a continuation event if your project is not finished).
For further details or support with the application process, please contact the NAC-WG.
Dardel-GPU is the accelerated partition based on AMD’s Instinct MI250X GPU of the Cray EX system from Hewlett Packard Enterprise. It has an accompanying Lustre storage system.
The nodes are interconnected using Slingshot HPC Ethernet.
Bianca is a research system dedicated to analysing sensitive personal data, or other types of sensitive data.
Bianca provides 4480 cores in the form of 204 dual CPU (Intel Xeon E5-2630 v3) Huawei XH620 V3 nodes with 128GB memory, 75 fat nodes with 256 GB of memory, 15 nodes with 512 GB of memory, ten nodes with two NVIDIA A100 40GB GPUs each.
Project storage attached to Alvis and Vera, dedicated for AI/ML
Project storage attached to Alvis and Vera, dedicated for AI/ML
Mimer is an all-flashed based storage system based on as solution from WEKA IO.
It consists of an 0.6 PB all-flash tier and a 7 PB Ceph based bulk storage tier (with spinning disk).
Swedish share of project storage attached to LUMI-C and LUMI-G
Project storage for NAISS allocations on LUMI.
Storage is applied for using TB-hours. Flash storage, LUMI-F, is accounted at ten times the TB-hour rate, i.e. use of 1TB of Flash storage for one hour costs 10 TB-hours. Lustre storage, LUMI-P, is accounted at the TB-hour rate. CEPH object storage, LUMI-O, is accounted at ½ the TB-hour rate, i.e. use of 1TB of CEPH storage for one hour costs 0.5 TB-hours. The total size of the Swedish part LUMI storage system is 35 412 000 TB-hours.
Project disk storage for Arrhenius GPU and Arrhenius CPU
Project disk storage for projects with compute allocations on Arrhenius GPU and/or Arrhenius CPU.
Using the Lustre parallel filesystem, it is designed for fast access from Arrhenius GPU and Arrhenius CPU. File contents are stored on spinning disks. The total size of Arrhenius Disk is 25 PB. There is no backup of Arrhenius Disk.
Project flash storage for Arrhenius GPU and Arrhenius CPU
Project flash storage for projects with compute allocations on Arrhenius GPU and/or Arrhenius CPU.
Using the Lustre parallel filesystem, it is designed for fast access from Arrhenius GPU and Arrhenius CPU. File contents are stored on NVMe SSDs, allowing for a higher small random IOPS compared to Arrhenius Disk. The total size of Arrhenius Flash is 2 PB. There is no backup of Arrhenius Flash.
Arrhenius Flash is only for projects that benefit significantly from the higher small random IOPS.
Project storage attached to Arrhenius SENS Compute
Project storage for projects with compute allocations on the NAISS compute system for sensitive data, Arrhenius SENS Compute.
This resource should only be requested in a proposal together with a request for Arrhenius SENS Compute. NAISS SENS Storage and Compute will only be allocated together in a project and they can only be used together within that project..
The total size of Arrhenius SENS Storage is 10 PB. There is no backup on Arrhenius SENS Storage.
Project storage for NAISS as well as LiU Local projects with compute allocations on resources hosted by NSC.
Centre Storage @ NSC is designed for fast access from compute resources at NSC. It consists of one IBM ESS GL6S building block and one IBM ESS 5000 SC4 building block.
In total there are 946 spinning hard disks and a small number of NVRAM devices and SSDs which act as a cache to speed up small writes. The total disk space that is usable for storing files is approximately 6.9 PiB.
Project storage for NAISS as well as PDC projects with compute allocations on resources hosted by PDC.
Klemming is designed for fast access from compute resources at PCD. It uses the Lustre parallel file system, which is optimized for handling data from many clients at the same time. The total size of Klemming is 12 PB.
National storage infrastructure for large-scale research data
Swestore is a Research Data Storage Infrastructure, intended for active research data and operated by the National Academic Infrastructure for Supercomputing in Sweden, NAISS,
The storage resources provided by Swestore are made available for free for academic research funded by VR and Formas through open calls such that the best Swedish research is supported and new research is facilitated.
The purpose of Swestore allocations, granted by National Allocations Committee (NAC), is to provide large scale data storage for “live” or “working” research data, also known as active research data.
Arrhenius Persistent Compute and Data (PCD) is a national cloud service provided by NAISS that gives researchers in Sweden flexible access to computing resources for their work. It is designed to support projects that benefit from adaptable environments, such as developing new tools, testing ideas, or running data-driven applications. PCD complements traditional high-performance computing by offering a more customizable and accessible platform, helping researchers explore and innovate across a wide range of disciplines.
Swedish Science Cloud provides Infrastructure as a Service; IaaS
Swedish Science Cloud (SSC) is a geographically distributed OpenStack cloud Infrastructure as a Service (IaaS), intended for Swedish academic research provided by NAISS.
It is available free of charge to researchers at Swedish higher education institutions through open application procedures.
The SSC resources are not meant to be a replacement for NAISS supercomputing resources (HPC clusters). Rather, it should be seen as a complement, offering advanced functionality to users who need more flexible access to resources (for example more control over the operating systems and software environments), want to develop software as a service, or want to explore recent technology such as for “Big Data” (e.g. Apache Hadoop/Spark) or IoT applications.
Examples of applications that must run in the normal HPC clusters since the SSC isn't built for those purposes.
AI/LLM/Deep Learning and other applications relying on GPU resources cannot be run in the SSC due to the limited access to GPU's.
Applications assuming high performance or large volumes of storage, network, or cores.
Benchmarking will not perform well within SSC due to performance limitations.
SSC isn't classified for Sensitive data, those project MUST use Cluster system for sensitive data within NAISS.
Other National Resources
Resources funded outside of NAISS but which are nationally available, in some cases under special conditions. See conditions for access under each resource.
KAW financed resource dedicated for AI/ML research
Berzelius Ampere is an NVIDIA® SuperPOD consisting of 94 NVIDIA® DGX-A100 compute nodes supplied by
Atos/Eviden and 8 CPU nodes also supplied by Eviden. The original 60 "thin" DGX-A100 nodes are each equipped with 8 NVIDIA® A100 Tensor Core GPUs, 2 AMD Epyc™ 7742 CPUs, 1 TB RAM and 15 TB of local NVMe SSD storage. The A100 GPUs have 40 GB on-board HBM2 VRAM. The 34 newer DGX-A100 nodes "fat" are each equipped with 8 NVIDIA® A100 Tensor
Core GPUs, 2 AMD Epyc™ 7742 CPUs, 2 TB RAM and 30 TB of local NVMe SSD storage. The A100 GPUs have 80 GB on-board HBM2 VRAM. The CPU nodes are each equipped with 2 AMD Epyc™ 9534 CPUs, 1.1 TB RAM and 6.4 TB of local NVMe SSD storage.
Fast compute interconnect is provided via 8x NVIDIA® Mellanox® HDR per DGX
connected in a non-blocking fat-tree topology. In addition, every node is
equipped with NVIDIA® Mellanox® HDR dedicated storage interconnect.
All nodes have a local disk where applications can store temporary files. The
size of this disk (available to jobs as `/scratch/local`) is 15 TB on "thin"
nodes, 30 TB on "fat" nodes, and 6.4 TB on CPU nodes, and is shared between
all jobs using the node.
KAW financed resource dedicated for AI/ML research
The latest phase of the Berzelius service is Berzelius Hopper. Berzelius Hopper consist of 16 NVIDIA® DGX-H200 compute nodes supplied by Eviden and 8 CPU nodes also supplied by Eviden.
The DGX H200 are equiped with 8 NVIDIA® H200 141GB GPUs, 2 Intel® 8480C CPUs, and 2.1 TB RAM. The CPU nodes are each equipped with 2 AMD Epyc™ 9534 CPUs, 1.1 TB RAM and 6.4 TB of local NVMe SSD storage. The DGX H200 nodes are connected to a fast interconnect with 8x NVIDIA® Mellanox® NDR per DGX in a non-blocking fat-tree topology. This is a separate interconnect from that which connects the DGX A100 nodes in Berzelius Ampere.
All nodes have a local disk where applications can store temporary files. The
size of this disk (available to jobs as `/scratch/local`) is 30 TB on H200 nodes, and 6.4 TB
on CPU nodes, and is shared between all jobs using the node.
Berzelius Hopper is accessed through a new set of login nodes separate from those in the original Berzelius installation and also has new servers for other supporting tasks.
Berzelius Hopper is currently in a testpilot phase before being released for new projects in SUPR.
SciLifeLab’s national OMERO service for collaborative bioimage data management and visualisation. Managed in collaboration between Chalmers e-Commons and SciLifeLab Data Centre.
Shared, central storage accessible from all Berzelius Ampere and Berzelius Hopper compute and login nodes is provided by a storage cluster from VAST Data consisting of 8 CBoxes and 3 DBoxes using an NVMe-oF architecture. The storage servers are connected end-to-end to the GPUs using a high bandwidth interconnect separate from the East-West compute interconnect. The installed physical storage capacity is 3 PB, but due to compression and deduplication this will be higher in practice.
NSC centre storage (as available on Tetralith) is not accessible on Berzelius.
SciLifeLab DDLS funded storage resource. Managed in collaboration between UPPMAX and SciLifeLab Data Centre.
S3 object storage.
Local and Regional Resources
Resources financed by individual universities or in regional collaborations between universities. Access is often limited to employees of the universities where the resources are located. See conditions for access under each resource.
COSMOS represent a significant increase in computational capacity and will offer access to modern hardware including GPUs. Through the LUNARC Desktop new and existing users will be able to draw upon the benefits of high performance computing (HPC) while not having to be burdened by the intricacies of HPC utilisation. At the same time, users proficient in HPC usage will still be able to make use of the computational power represented in the interconnected nodes of COSMOS.
COSMOS consists out of 182 compute nodes funded by Lund university. Each node has two AMD 7413 processors (Milan), offering 48 compute cores per node. The nodes have 256 GB ram installed. In addition to the CPU nodes there are also 6 NVIDIA A100 nodes and 6 NVIDIA A40 GPU Nodes. For more specs see below.
COSMOS also features Intel partitions with Intel processors (Caskade Lake), offering 32 compute cores each. There are 22 CPU nodes, 5 nodes with NVIDIA A40 GPUs and four nodes with A100 GPUs within the Intel partitions.
Local resource for Linköping University researchers
Sigma, sigma.nsc.liu.se, runs a Rocky Linux 9 version of the NSC Cluster Software Environment. This means that most things are very familiar to Gamma users.
You still use Slurm (e.g sbatch, interactive, ...) to submit your jobs. ThinLinc is available on the login node. Applications will still be selected using "module".
All Sigma compute nodes have 32 CPU cores. There is 104 "thin" nodes with 96 GiB of primary memory (RAM) and 4 "fat" nodes with 384 GiB. Each compute node have a local SSD disk where applications can store temporary files (approximately 200GB per node).
All Sigma nodes are interconnected with a 100 Gbps Intel Omni-Path network which is also used to connect the existing storage. The Omni-Path network work in a similar way to the FDR Infiniband network in Gamma (e.g still a fat-tree topology).
Sigma have a capacity that exceeds the current computing capacity of Gamma. Sigma was made available to users on August 23, 2018.
Dardel is a Cray EX system from Hewlett Packard Enterprise, based on AMD EPYC processors with an accompanying Lustre storage system.
The nodes are interconnected using Slingshot HPC Ethernet.
Dardel-GH is the accelerated partition based on nVIDIA GraceHopper GPUs. It has an accompanying Lustre storage system. The nodes are interconnected using Slingshot HPC Ethernet.
A new high-performance GPU resource is now accessible on the Dardel supercomputer at PDC, providing 20,000 GPU hours per month. Eight additional nodes have been purchased, each equipped with four NVIDIA Grace Hopper (H100) GPUs and four NVIDIA Grace ARM CPUs—totalling 32 Grace Hopper Superchips. This state-of-the-art platform combines the performance of the Hopper GPU with the energy-efficient Grace ARM CPU, designed for demanding AI and high-performance computing (HPC) workloads.
The Grace Hopper resource is exclusively available to projects under the Medium compute category, with priority given to users running codes not currently supported on Dardel's AMD GPUs. If there is a need to run unique or specialized software, this new system offers a significant opportunity for improved performance.
Important:
For users with an existing Medium project on Dardel with a need to use dardel-GH resource, submit a continuation of your existing Medium project (you can submit a continuation event if your project is not finished).
For further details or support with the application process, please contact the NAC-WG.
This UU-funded resource was installed in late 2024.
It features standard compute nodes with 48 cores and 768 GB per node, many nodes with Nvidia T4 or L40s GPUs, and some nodes with Nvidia H100 GPUs.
Project storage attached to Alvis and Vera, dedicated for AI/ML
Project storage attached to Alvis and Vera, dedicated for AI/ML
Mimer is an all-flashed based storage system based on as solution from WEKA IO.
It consists of an 0.6 PB all-flash tier and a 7 PB Ceph based bulk storage tier (with spinning disk).
Project storage for NAISS as well as LiU Local projects with compute allocations on resources hosted by NSC.
Centre Storage @ NSC is designed for fast access from compute resources at NSC. It consists of one IBM ESS GL6S building block and one IBM ESS 5000 SC4 building block.
In total there are 946 spinning hard disks and a small number of NVRAM devices and SSDs which act as a cache to speed up small writes. The total disk space that is usable for storing files is approximately 6.9 PiB.
Project storage for NAISS as well as PDC projects with compute allocations on resources hosted by PDC.
Klemming is designed for fast access from compute resources at PCD. It uses the Lustre parallel file system, which is optimized for handling data from many clients at the same time. The total size of Klemming is 12 PB.
