Team FAUltier from Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU) achieved an outstanding result at the ASC26 Student Cluster Competition, held from May 16–20, 2026, at Wuxi University in Jiangsu Province, China. The team was ranked as the highest-performing international and non-Mainland China team in the competition and secured 12th place overall among the world’s best student high-performance computing (HPC) teams. In addition, the team’s advisor received the competition’s Most Popular Advisor Award.
After successfully mastering the online preliminary round in February, where approximately 300 teams from around the world competed, FAUltier advanced to the finals as one of only 25 selected teams. The final round took place on-site in China and challenged participants to assemble and set up an HPC cluster with the provided hardware, followed by completing six different tasks over the span of two days, ranging from HPL benchmarking to working with simulation and AI.
One of the World’s Largest Student Supercomputing Competitions
Alongside the Student Cluster Competition (SCC) in the United States and the ISC Student Cluster Competition in Germany, the ASC Student Supercomputer Challenge is one of the three major international student cluster competitions and the largest supercomputing hackathon worldwide. Organized by the Asia Supercomputing Community, the competition has been held annually since 2012.
FAU has participated in the ASC competition six times, from 2017 to 2019 and again since 2024, often sending the team who gained experience during the SCC in the year before; for example, this year’s team was assembled for the IndySCC at SC25 in Saint Louis, USA.
Intensive Preparation and Cluster Design
The preparation phase began with the tasks for the preliminary round. There, the team had to optimize test cases for “Embodied World Model” (official name: “UnifoLM World-Model-Action“) and the numerical relativity program AMSS_NCKU, run the renowned HPL and HPCG benchmarks, and plan the setup of a cluster that would perform optimally under the competition constraints of a maximum total power consumption of 5 kW and a per-node limit of 2 kW.
A major focus during the preparation phase was cluster architecture and performance optimization. Using the hardware specifications provided by the ASC Committee, particularly the AMD EPYC 9755 CPU servers, the team evaluated different cluster configurations and performance characteristics. Extensive testing was carried out on local university clusters to gain experience with networking setups, node communication, and system stability. Different software configurations were explored to identify potential bottlenecks and ensure a robust and reliable cluster deployment during the competition.
Since success in the competition depends not only on raw performance but also on maintaining operation within the strict 5 kW power budget, the team experimented with various tools and system-level settings to precisely control CPU power consumption. This included testing different power limits, frequency settings, and performance profiles to find the optimal balance between computational performance and energy efficiency. These experiments provided valuable insights into how the cluster could be tuned dynamically to maximize benchmark and application performance while remaining within the competition’s power constraints.
The preparation efforts were further strengthened by the support of Team FAUsion, whose considerable experience and technical expertise proved invaluable in addressing challenges related to cluster design and networking. We also gratefully acknowledge Regionales Rechenzentrum Erlangen (RRZE) for providing the laptops used during the competition and the Department Informatik and Erlangen National High Performance Computing Center (NHR@FAU) for travel support. Their backing significantly contributed to our team’s ability to participate effectively.
Outstanding Benchmark Performance
The competition itself spanned four intensive days. The first two were spent building a small cluster with the hardware provided by the ASC committee. Finishing the cluster after one day, the team was able to test all applications one last time on the competition hardware before entering the final stage of the competition and figuring out the optimal power limit settings. On day three, the team was tasked with running HPL and HPCG benchmarks. These ran smoothly and showed by far the best results from all CPU clusters in the competition, even beating some GPU clusters, partially due to the team using more InfiniBand cards than the other teams and disabling individual cores to increase the boost clock for the memory-bound HPCG benchmark.
Tackling Scientific Simulations and AI Workloads
Following the benchmarking phase, teams were tasked with running Embodied World Model, an artificial intelligence application designed to predict physical processes for robotics training. At the same time, the competition’s international group challenge required teams from different countries to collaborate on optimizing and executing the ICON atmospheric forecasting tool developed by the DKRZ in Hamburg. FAUltier would like to explicitly thank Tsinghua University for very cooperative teamwork and a great time working together.
The final competition day presented perhaps the most demanding scenario, as three tasks had to be completed at once. QiboTN, a tool for simulating quantum computers, uses tensor networks instead of state vectors and therefore saves a lot of memory; AMSS_NCKU, a black hole physics simulation and the mystery application LeWorldModel, another AI model for predicting physics behavior. However, the greatest challenge was not optimizing the tasks, but efficiently distributing the sparce compute resources to finish all tasks before deadline, which was achieved with only 30 seconds left on the clock.
Overcoming GPU Software Challenges
One of the most significant technical obstacles involved the GPU-based Embodied World Model task. Additionally to the fact that not all dependencies were available in the right versions with ROCm, which was needed because the provided GPUs were manufactured by AMD, the provided source code we got on the competition had an accuracy problem with ROCm. Therefore, a lot of time was spent on finding out about the exact cause of the issue and solving it, but at last there was a patch found by working together with different teams and exchanging information and search results to that issue.
Valuable Experience Beyond Competition Results
FAU has a long tradition of participating in international student cluster competitions. Future competitions, as well as the exchange of knowledge and experience with former teams, will continue to provide students with valuable learning opportunities. This year’s advisor, Melanie, previously competed as a student team member at SC24 and SC25 herself.
They will also help expand FAU’s presence within the international high-performance computing community and create valuable networking opportunities with universities from around the world. A long-term goal of these competitions is to establish a sustainable student cluster team at FAU. By maintaining continuous involvement in these events, FAU can strengthen its expertise in high-performance computing while ensuring that future generations of students benefit from the experience gained by previous teams. Continued participation will also help attract and inspire new students to engage with high-performance computing and related research fields.
The success of Team FAUltier at ASC26 represents another important milestone in FAU’s growing presence within the international supercomputing community.
This article was written by the members of team FAUltier in collaboration with the editorial staff of NHR@FAU.
2/4
3/4
4/4