Helmholtz GPU Hackathon 2025 – Intensive Hacking towards JUPITER

From April 1 to April 11, 2025, Jülich Supercomputing Centre (JSC) of Forschungszentrum Jülich hosted the Helmholtz GPU Hackathon 2025, bringing together 52 researchers, developers, and data scientists from across Helmholtz institutes and beyond. The hackathon, now firmly established as a highlight of the Helmholtz calendar, remains centered around optimizing and accelerating complex HPC applications with AI codes now blending into the mix. Preparation for the JUPITER exascale machine was a key focus this time around.

The event was organised in cooperation with Helmholtz HIDA, NVIDIA, and the OpenACC Organization, as part of the Open Hackathons program, and unfolded in two key phases: a virtual kickoff followed by a highly collaborative on-site week in Jülich.

Hack, Collaborate, Accelerate: A Hackathon-Week at JSC

The hackathon officially began with two virtual onboarding sessions held on March 25 and April 1 as a virtual introduction meeting with all teams and mentors. These prior meetings offered a chance for all participants to connect and familiarise themselves with each other and prepare their codes for what was coming next. This set the stage perfectly for the on-site hackathon week.

Once the teams were set, the real action moved on-site to JSC, where participants gathered in person from April 8 to April 11 for four days of intensive, hands-on GPU development.
Each day focused on hands-on teamwork to advance the code, culminating in brief team presentations to share progress and insights. The close cooperation between mentors and other participants made the event not only highly productive for the team but also enriching for the open scientific community.

 “Together with mentors, our team investigated the potential bottlenecks and different porting approaches, including NVIDIA CUDA and OpenACC, that made two modules with expensive functions run 1.5 times faster,” shared Dr. Martin Kühn, Research Group leader and main contributor to the MEmilio software framework. “We left the hackathon with more insights into our codebase and a lot of ideas for future development.”

From Galaxies to Oceans: Scientific Research Accelerated on JEDI

7 Teams across HPC and AI domains spent the time exploring how to bring their performance up to top speed, especially to the NVIDIA GH200 superchip powering JUPITER, the first exascale system in Europe. While the JUPITER construction site was still busy a mere 50 m from the hackathon venue, first hardware of the system could be used through the JEDI early access machine – still the most energy-efficient supercomputer in the world.

Among the returning teams was team GALÆXI, exploring the performance of FLEXI, an open-source high-order numerical framework for solving partial differential equations (PDEs), with a special focus on Computational Fluid Dynamics.

Having attended the Helmholtz GPU Hackathon in the previous two years, the team worked out an approach to use the Hackathon week as a springboard for a new phase in the development of GALÆXI, the GPU-ported version of FLEXI. This year, they spent the week porting key FLEXI features to GALÆXI and optimizing the core solver for GH200. A dedicated focus was the JUPITER Research and Early Access Program (JUREAP), in which the GALAEXI team participates.

“At the Hackathon, we achieved approximately a 30% speedup in the most expensive parts of our code, and an 8.5× speedup on Grace-Hopper GH200 over A100. We are getting close to running our simulation on basically every supercomputer in Europe and running really fast with all the features we need.”

Another returning team, first at the Hackathon in 2018, is McStas+McXtrace. The team worked on the combination of instrument- and experiment-simulation tools for neutron- and X-ray scattering at large scale research facilities, like the European Spallation Source. McStas and McXtrace, the two simulation tools, were ported to GPUs already in the past years, but the team sought to gain more out of their performance. Certain algorithms presented themselves as slow for big problems and would benefit from more targeted GPU optimisation.

“We thought our GPU performance had plateaued and hoped that the Open Hackathon would inspire us to improve what we could do,” described Dr. Tobias Weber from Institut Laue-Langevin in Grenoble, France. “By changing the memory structure and the number of kernels, our simulation now runs 2 times faster and we already sketched further efficiency improvements.”

Team HybridNEMO, with researchers from European Centre for Medium-Range Weather Forecasts (ECMWF) and Barcelona Supercomputing Center (BSC), worked on optimizing the NEMO (Nucleus for European Modelling of the Ocean) ocean model for efficient execution on GPU architectures. NEMO is a powerful and open-source framework, used to simulate the ocean’s dynamics, sea ice, and biogeochemistry. The team conducted extensive previous work on optimizing NEMO and brought the framework to the hackathon to improve its scalability on GPUs, especially JUPITER and the GH200 superchip. The team is also part of JUREAP.

Other participating teams – team CMS Digi Morphing representing CMS, CERN, and VUB, team CP2K from HZDR/CASUS, and team Computer Says Snow from ECMWF – had a chance to explore JEDI capabilities over the duration of the hackathon. The teams realized immediate performance improvements and future-proofed their applications and paved the way for more efficient operations in the long run.

What’s next?

Leaving the hackathon, the teams not only enjoyed the achievements of the past week, but also gathered ideas for the next steps. The tight collaboration between technical experts and computational scientists resulted in new connections for potential future project development.

A big thanks to the JSC team, community mentors, and all participants for making 2025’s Hackathon a leap forward in GPU-empowered research!