Giga Computing × National Taiwan University: SC25 Student Cluster Competition Champions

The combination of High-Performance Computing and AI has become the new frontline of world computing infrastructure. Recognizing this shift, academic institutions around the world are investing significant resources to cultivate AI/HPC talent capable of system integration and real-time decision-making, the very skills on display at the Student Cluster Competition (SCC). Held annually since 2007 alongside the Supercomputing Conference, the SCC is a showcase of the world's most prestigious HPC talent.

In 2025, a team from National Taiwan University, led by faculty advisor Prof. Chun-Yi Lee, competed in the SCC for the very first time. With full hardware support and technical collaboration from Giga Computing, the NTU team impressed judges with their system architecture design, performance tuning, and workload optimization capabilities. In a grueling 48-hour challenge, they outperformed seven of the world's top teams to claim the overall championship, a landmark achievement for Taiwan on the global HPC stage.

"The NTU team winning the overall championship in their very first SC Student Cluster Competition proves that Taiwanese students possess world-class capabilities in high-performance computing and artificial intelligence," stated the team's faculty advisor, Prof. Chun-Yi Lee.

At the heart of the SCC competition is the requirement for each team to complete a series of tasks spanning scientific simulation, AI workloads, and system benchmarks within a strict 10,000W power budget. Maximizing overall cluster performance within this fixed power envelope is the SCC's defining challenge, and it was the central focus of the NTU team's training ahead of the event.

This year's competition tasks included HPL/HPL-MxP benchmarks, the Exascale Climate Emulator, the Structural Simulation Toolkit, and a Mystery Application (ACTS) revealed only on the final day. These tasks span atmospheric simulation, computer architecture design, and AI inference, requiring teams to make system architecture, performance tuning, and resource allocation decisions in a very short timeframe.

Since the organizers provided only twelve C14/C20 outlets on-site, teams had to precisely plan power distribution and thermal management across multiple nodes, a test not only of technical depth, but of real-time engineering judgment under pressure.

As the hardware partner for this competition, Giga Computing provided key support for the NTU team, ensuring they had the closest possible experience to a real world supercomputing environment, from preparation through competition day.

• Two months before the competition, Giga Computing opened its lab to the team for comprehensive system testing, with full BMC and server access provided throughout.
• Deployed four GIGABYTE G494-ZB4-AAP2 GPU servers equipped with AMD EPYC 9655 processors, NVIDIA H200 GPUs, NVIDIA ConnectX-7 high-speed NICs, Micron DDR5 memory, and Solidigm PCIe 5.0 SSDs, providing the computational foundation for high-density computing, generative AI inference, distributed HPC, and I/O-intensive workloads.
• Supplied custom-designed power cables and provided hands-on engineering guidance on server fan architecture, helping the team pinpoint the optimal settings for reducing power draw without risking CPU or GPU overheating.
• Offered real-time on-site support to confirm maximum cable power ratings at critical moments, giving the team the confidence to make fast, informed decisions.

This collaboration was facilitated through an introduction by the National Center for High-performance Computing (NCHC) under the National Applied Research Laboratories, which served as an important bridge connecting NTU and Giga Computing for this industry-academia partnership.

Scaling from Three Nodes to Four

During preparation, the team discovered that by taking a portion of the CPU cores offline and powering them down, a single server's idle power consumption could be reduced from approximately 700W down to around 200W. This discovery freed up enough power to upgrade from the originally planned three node architecture to four nodes, ultimately settling on an optimal two GPU, two CPU configuration.

Precision Power Management and the Full-Speed Decision

With limited outlets available, the team allocated power using a "4/4/2/1" distribution strategy, directing the bulk of capacity to the GPU nodes. As the competition entered its final stretch, they determined that playing it safe wouldn’t allow them to complete all tasks within the time limit. After a quick check to confirm the cables' maximum power rating, they made the call to push all servers to full speed, one of the most decisive moments of the entire competition.

Hardware-Level Performance Optimization

• GPU tasks: Took full advantage of the 8-GPU-per-node architecture, maximizing the much faster intra-node communication for a decisive competitive edge.
• CPU tasks: With power headroom available, the team raised fan speeds to hold CPU temperatures around 90°C, keeping the processors locked at maximum boost frequency.
• Mystery Application (ACTS): Identifying the task as CPU-intensive and scored on execution count, the team set aside extensive code modifications and concentrated resources on maximizing the number of runs, a competition-oriented judgment that prioritized scoring impact over a purely academic approach.

The NTU team was composed of six students from different departments including Computer Science, Electrical Engineering, Mechanical Engineering, and Ocean Engineering, demonstrating the depth of interdisciplinary collaboration. The team identified performance profiling as their greatest technical takeaway from the competition:

• Deeply analyzing the underlying computational behavior of each task to precisely identify true performance bottlenecks, determining whether they were memory-bound or compute-bound before deciding on an optimization direction.
• Making holistic trade-off decisions balancing power consumption, hardware constraints, task characteristics, and time pressure, rather than pursuing single-point maximum performance.

Additionally, the team had previously competed in the 2025 ASC Student Supercomputer Challenge, winning the Team Competition Award and First Prize, accumulating rich international competition experience that laid a solid foundation for their SC debut.

"This experience tested students' system tuning and problem-solving abilities under extreme conditions. We also thank our industry partners for providing the latest technology, giving students the closest possible experience to operating a real supercomputing center," said Faculty advisor Prof. Chun-Yi Lee after the victory.

NCHC Deputy Director Chih-Huang Hsiao stated: "The NTU team's outstanding performance in SCC25 fully demonstrates the results of NCHC's long-term efforts to promote AI/HPC education domestically. We are committed to ensuring that student research teams have access to top-tier computing resources and gain valuable experience through hands-on competition. Going forward, NCHC will continue to play the role of bridging industry resources to jointly cultivate a new generation of HPC and AI talent with a global perspective."

"We are incredibly proud to see the NTU team win the overall championship in their first SCC appearance. Giga Computing not only provided a high-performance server platform but also participated deeply in system optimization and technical support,” said Chen Ming-Jiang, Vice General Manager of Giga Computing. “This is the best demonstration of our commitment to the value of industry-academia collaboration. Going forward, we will continue to empower student research teams with technology, allowing Taiwan's deep expertise in server system integration to shine on the world stage."

This deep collaboration between Giga Computing and National Taiwan University points to a repeatable model for building HPC talent: putting students alongside production-grade hardware and the engineers who run it, so that classroom theory meets the real constraints of a working data center. As AI and HPC workloads continue to converge, that hands-on grounding is what will set apart the engineers who go on to design and operate the world's next computing infrastructure.