Format: Hybrid training
Dates: 14–15 September 2026
Location: CSC Training facilities, Keilaranta 14, Espoo and online.
Target system: Roihu, NVIDIA Grace Hopper GH200 GPUs
Event overview
Roihu is CSC’s new national supercomputer based on NVIDIA Grace Hopper GH200 nodes. The system combines NVIDIA Hopper GPUs with Grace Arm CPUs and provides a programming environment that differs in important ways from previous CSC GPU systems such as Mahti.
This event is intended for researchers and developers who want to understand how to use Roihu efficiently, especially for scientific and high-performance computing applications.
The training is suitable for a broad audience interested in GPU programming and performance on Roihu. The training will be delivered by experts from NVIDIA.
Right after the training, there is a hackathon is intended for code developer teams with a concrete porting, profiling, or optimization goal.
Link to hackathon page/part.
Location
On site
CSC Training facilities
Keilaranta 14
02150 Espoo
Finland
On-site participation is recommended especially for participants who will continue to the hackathon part of the event. Hackathon participants will be prioritized for on-site attendance.
Remote Participation
The training part is planned as a hybrid event. A remote participation link will be provided to registered participants closer to the event.
Remote participation is available for the training part only. Interaction with teachers during the hands-on parts may be more limited for remote participants.
Who should attend?
The training is aimed at Finnish researchers, students, application developers, and CSC users interested in running GPU applications on Roihu.
The course is especially useful for participants who:
- Develop scientific or high-performance computing applications.
- Have previous experience with GPU programming or GPU-accelerated applications.
- Want to learn about NVIDIA Grace Hopper architecture and its performance implications.
- Want to profile and optimize applications on NVIDIA GPUs.
Prerequisites
Participants are expected to have:
- Basic Linux command-line skills.
- Basic experience with HPC systems and batch jobs.
- Some programming experience in C, C++, Fortran, Python, or another scientific programming language.
- Familiarity with GPU programming concepts.
- Access to Roihu
Topics
The training will cover topics such as:
- Overview of Roihu and the NVIDIA Grace Hopper GH200 architecture.
- Differences between Roihu GH200 and previous CSC GPU systems such as Mahti A100.
- Grace CPU and Hopper GPU architecture.
- Memory hierarchy and unified memory considerations.
- Programming models for NVIDIA GPUs:
- CUDA
- OpenMP offload
- OpenACC
- GPU-accelerated libraries
- Multi-GPU and multi-node considerations.
- NVIDIA HPC SDK and compiler environment.
- CUDA versions and CUDA 13 considerations.
- Profiling and performance analysis with Nsight Systems and Nsight Compute.
- Practical performance considerations for scientific applications.
Learning outcomes
After the training, participants should be able to:
- Describe the main architectural features of NVIDIA Grace Hopper GH200 nodes.
- Understand how Roihu differs from previous CSC GPU systems such as Mahti.
- Identify important performance considerations for GH200 applications.
- Understand the role of unified memory and CPU-GPU data movement on GH200.
- Compile and run GPU applications on Roihu.
- Use NVIDIA profiling tools to identify performance bottlenecks.
- Prepare their own code for further optimization during the hackathon or future work.
Agenda
The agenda is preliminary and may still be adjusted.
The training will be delivered by experts from NVIDIA.
14 September 2026: Roihu GH200 training, part 1
13:00–17:00
- Welcome and practical information.
- Introduction to Roihu and NVIDIA Grace Hopper GH200.
- Overview of GPU computing on Roihu:
- Comparison with previous CSC GPU systems.
- Grace Hopper architecture overview:
- Grace Arm CPU.
- Hopper GPU architecture.
- CPU-GPU connectivity.
- Memory hierarchy.
- Programming the NVIDIA platform:
- CUDA ecosystem.
- GPU-accelerated libraries.
- Applications and frameworks.
- Standard languages and directives.
- CUDA, OpenMP offload, and OpenACC.
- Compilers and developer tools.
15 September 2026: Roihu GH200 training, part 2 and hands-on lab
09:00–12:00
-
Grace Hopper programming considerations:
- How the architecture affects code optimization.
- Caches and local/shared memory.
- Unified memory: when and how to use it.
- Keeping data close to the GPU.
- NUMA considerations.
-
NVIDIA software environment on Roihu:
- NVIDIA HPC SDK.
- CUDA versions and CUDA 13 considerations.
-
Multi-GPU and multi-node considerations:
- NVLink.
- Library support.
- MPI and NCCL considerations.
- Profiling multi-node applications.
12:00–13:00
Lunch break.
13:00–17:00
Hands-on lab.
The hands-on session will focus on profiling and performance analysis. The exact exercises are preliminary, but the session is expected to include:
- Running example GPU applications on Roihu.
- Profiling with NVIDIA Nsight Systems.
- Kernel-level profiling with NVIDIA Nsight Compute.
- CPU-GPU data movement.
- Investigating performance bottlenecks.
- Optional experimentation with CUDA, OpenACC, and/or OpenMP offload examples.
The hands-on lab also serves as preparation for teams participating in the Roihu GH200 hackathon later in the week.
Registration
The training is a hybrid event. On-site participants should register in advance. Remote participation details will be provided to registered participants before the event.
All on-site participants for the training should register by 15 August 2026. Online registrations is possible until 7 September 2026.
Organizer
CSC – IT Center for Science
NVIDIA
Attendance fee
Free for all Finnish institutions members with access to Roihu.
Comments and questions
For questions about the event, please contact: