Efficient use of the Grace Hopper GPU nodes on Dardel and Arrhenius¶

This course is a collaboration between NAISS and PDC. The lectures will be provided by Nvidia and Hewlett Packard Enterprise (HPE)

30 - 31 March 2026

This training course will discuss the efficient use of the Grace Hopper GPU nodes on NAISS’ Dardel and NAISS’ forthcoming Arrhenius system.

Dardel is NAISS’ current flagship service hosted at PDC in KTH. Arrhenius which is currently being installed will be NAISS’ forthcoming powerful supercomputer that is build together with EuroHPC. Both systems are provided by Hewlett Packard Enterprise (HPE) and include NVIDIA Grace Hopper GPU nodes. The course will delivered by consultants from NVIDIA and HPE.

This course is expected to be of particular interest for research groups developing scientific software to be deployed on either Dardel or Arrhenius or both.

This course is intended for experienced users of HPC.

Location¶

On site¶

E-huset, Room E32
Lindstedtsvägen 3
Stockholm

which is located on the KTH campus.

Map showing the location of the venue (on google.com)

Remote¶

Closer to the event a zoom link will be provided to registered participants. Inter action with teachers around the event and in particular during the hands-on parts on the second day for remote participants might be limited.

Schedule¶

• 30th March 2026: 9:00 - 17:00
• 31st March 2026: 9:00 - 17:00

Agenda¶

The course covers the following topics (preliminary agenda)

30 March 2026: Grace Hopper day¶

Taught by consultants from NVIDIA

9:00 - 12:30¶

• Introduction to GPU (30 min)

  • Basics of GPU computing
  • Comparison CPU/GPU
  • GPU hardware features

• Ways to accelerate applications (60 min)

  • CUDA ecosystem
  • Libraries
  • Applications and Frameworks
  • AI and HPC

• Programming the NVIDIA platform (30 minutes)

  • Standard languages, directives, CUDA
  • MultiGPU considerations
  • Compilers and developer tools

• Grace Hopper overview (90 minutes)

  • Grace - overview of Arm CPU
  • Hopper - features of the Hopper architecture
  • Grace Hopper - details

13:30 - 17:00¶

• Grace Hopper Programming Considerations (60 minutes)
• Porting to Arm (30 minutes)
• Nsight tools (30 minutes)

• GH and Applications Performance (60 minutes)

  • HPC apps, LLMs, extended memory

• Conclusion (30 minutes)

31 March 2026: HPE Cray day¶

Taught by consultants from Hewlett Packard Enterprise

9:00 - 9:55¶

• 9:00 Welcome
• 9:10 Cray architecture

• 9:40 Execise

• 9:55 Break

10:10 - 11:25¶

• 10:10 Cray Compilation Environment and GPU Parallel Programming Models

• 10:40 Exercises

• 11:25 Break

11:40 - 12:50¶

• 11:40 Understanding Cray MPI on Slingshot

• 12:10 Exercises

• 12:50 Lunch Break

13:50 - 15:00¶

• 13:50 Using NCCL with Slingshot

• 14:20 Exercise

• 15:00 Break

15:15 - 16:25¶

• 15:15 Debugging and profiling on GPUs

• 15:45 Exercise

• 16:25 Break

16:40 - 17:00¶

• 16:40 Open Questions & Answers / Finalise the exercises

17:00 Close of the day¶

There will be time in the schedule on 31 March for hand on exercises on NAISS’ Dardel system.

The performance and debugging tools mentioned above are available on Dardel. Their availability on Arrhenius is still under investigation.

Comments and questions¶

Comments and questions on these training events should be sent to NAISS using the support form in SUPR. Please select “Question about a training event” as “Problem or Questions Type”.

Registration¶

Please fill in the registration form:

Registration form for the course (on nextcloud.naiss.se)

Registration is set to close on 18 March 2026. Acceptance on the course will be announced no later than the 23 March 2026.