Skip to content

Docs CSC now features an automatic Finnish translation. Click here for more information.

Warning!

Puhti and Mahti computing services have been decommissioned and no new jobs are accepted or executed on its compute nodes. Puhti and Mahti login nodes and storage services are planned to remain available until 15 October 2026. Clean up unnecessary files and move any data you need to keep by 31 August 2026. See the Roihu data migration guide for instructions on transferring your data to Roihu.

Technical details about Puhti

Puhti retirement in Fall 2026

Puhti will be gradually decommissioned during summer 2026 and replaced by Roihu, CSC's next-generation supercomputer offering enhanced performance and capabilities.

Puhti computing services have been shut down 31 July 2026 at 12:00 EEST, and no new jobs are accepted or executed on its compute nodes.

Storage and login nodes are planned to remain accessible until midday 15 October 2026, but users are strongly encouraged to move any required data from the system without delay.

Learn more about Roihu

Compute

Puhti had a total of 682 CPU nodes, with a theoretical peak performance of 1,8 petaflops. Each node is equipped with two Intel Xeon processors, code name Cascade Lake, with 20 cores each running at 2,1 GHz. The cores supported AVX-512 vector instructions and VNNI instructions for AI inference workloads. The interconnect was based on Mellanox HDR InfiniBand. The nodes were connected with a 100 Gbps HDR100 link, and the topology was a fat tree with a blocking factor of approximately 2:1.

The Puhti AI artificial intelligence partition had a total of 80 GPU nodes with a total peak performance of 2,7 petaflops. Each node had two Intel Xeon processors, code name Cascade Lake, with 20 cores each running at 2,1 GHz. They also had four Nvidia Volta V100 GPUs with 32 GB of memory each. The nodes were equipped with 384 GB of main memory and 3,6 TB of fast local storage. This partition was engineered to allow GPU-intensive workloads to scale well across multiple nodes. The interconnect was based on a dual-rail HDR100 interconnect network connectivity providing 200 Gbps of aggregate bandwidth in a non-blocking fat-tree topology.

Nodes

Name Number of nodes Compute Cores Memory Local disk
M 484 Xeon Gold 6230 2 x 20 cores @ 2,1 GHz 192 GiB
M-IO 48 Xeon Gold 6230 2 x 20 cores @ 2,1 GHz 192 GiB 1490 GiB
L 92 Xeon Gold 6230 2 x 20 cores @ 2,1 GHz 384 GiB
L-IO 40 Xeon Gold 6230 2 x 20 cores @ 2,1 GHz 384 GiB 3600 GiB
XL 12 Xeon Gold 6230 2 x 20 cores @ 2,1 GHz 768 GiB 1490 GiB
BM-IO 6 Xeon Gold 6230 2 x 20 cores @ 2,1 GHz 1,5 TiB 5960 GiB
GPU 80 Xeon Gold 6230
Nvidia V100		 2 x 20 cores @ 2,1 GHz
4 GPUs connected with NVLink		 384 GiB
4 x 32 GB		 3600 GiB

In addition to the compute nodes above, Puhti has two login nodes with 40 cores and 2900 GiB local disk each.

Storage

Puhti has a 4.8 PB Lustre parallel storage system providing space for home, project and scratch storages.

Current Lustre configuration for Puhti is:

Storage area # OSTs # MDTs
home 24 4
projappl 24 4
scratch 24 4

Please see Lustre documentation for the terminology.

All the OSTs and MDTs are shared across the storage areas, thus the performance should be similar between them. The peak I/O performance in Puhti is around ~50 GB/s with 64 compute nodes under dedicated access to the system.