Xinnor, a leading provider of high-performance software-defined RAID solutions, today released a comprehensive case study detailing how its xiRAID technology, in partnership with MEGWARE, Celestica, and Phison, enabled Germany’s most powerful university-owned AI supercomputer at NHR@FAU to achieve the #3 position in the global IO500 benchmark rankings and #1 among Lustre-based solutions.
The Helma supercomputer at Friedrich-Alexander-Universität Erlangen-Nürnberg’s National High-Performance Computing Center (NHR@FAU) combines 192 dual-socket AMD EPYC 9554 “Genoa” compute nodes with 768 NVIDIA H100/H200 GPUs, ranking #51 on the June 2025 TOP500 list. The system supports cutting-edge research in deep learning for scientific applications, natural language processing with large language models, computer vision, reinforcement learning and robotics, and AI for engineering applications including digital twins and CFD optimization. The storage infrastructure, designed and built by MEGWARE using Celestica SC6100 systems with Phison Pascari drives and protected by Xinnor’s xiRAID Classic 4.2, delivered breakthrough performance metrics that established new benchmarks for high-availability NVMe storage in academic HPC environments.
The xiRAID-powered storage solution achieved extraordinary IO500 benchmark results:
- IOR Easy Read: 1,798.77 GiB/s
- IOR Easy Write: 811.33 GiB/s
- MDtest Easy Stat: 8,221.83 kIOPS
- Find Operations: 3,016.99 kIOPS
- Overall Bandwidth: 438.62 GiB/s
- Overall IOPS: 1,604.84 kIOPS
- Total IO500 score: 838.99
These results represent the highest IOps and bandwidth metrics ever achieved by a Lustre-based storage system in the IO500 rankings, demonstrating xiRAID’s ability to maximize hardware performance without compromising reliability.
The solution leverages a sophisticated half-rack design featuring 10x Celestica SC6100 Storage Bridge Bay Systems, each equipped with dual AMD EPYC 9454P processors, 384 GB DDR5 memory, and a combination of Phison Pascari PCIe 5.0 drives: 4x 6.4TB write-intensive drives for metadata and 20x 30.72TB read-intensive drives for data storage. The software stack combines xiRAID Classic 4.2 with Lustre 2.16.1 and integrated Corosync & Pacemaker clustering for high availability, running on AlmaLinux 9 operating system.
The xiRAID solution implements advanced fault tolerance through integrated Pacemaker and Corosync clustering. Each NVMe drive connects to both server nodes via separate PCIe lanes, with namespace splitting enabling simultaneous utilization from both nodes. This architecture ensures the system can survive multiple drive failures and complete server node failures with automatic failover and failback capabilities.
Prof. Gerhard Wellein, director of NHR@FAU: “The xiRAID-powered storage infrastructure has transformed our ability to support large-scale AI research. The elimination of I/O bottlenecks means our researchers can fully utilize Helma’s 768 GPUs for deep learning, quantum chemistry, and atomistic simulations. The high availability features ensure uninterrupted access to our petabyte-scale datasets, which is crucial for a national HPC center.”
Davide Villa, CRO, Xinnor: “Achieving the #3 global IO500 ranking and #1 position among Lustre solutions validates our commitment to delivering enterprise-grade performance without compromising on reliability or cost-effectiveness. The Helma deployment demonstrates how xiRAID can maximize the potential of cutting-edge hardware while maintaining the flexibility and economics of commodity components.”
Markus Hilger, HPC Engineer, MEGWARE: “The collaboration with Xinnor on the Helma project showcases the power of innovative software-defined storage. By combining xiRAID’s advanced RAID capabilities with our system integration expertise, we are able to build highly available storage solutions with performance very close to the hardware limits and no bottlenecks.”
Albert Kang, Sr. BD Director, Phison: “Our Pascari PCIe 5.0 drives were specifically chosen for this deployment to provide the perfect balance of performance and endurance for AI workloads. The integration with xiRAID’s software RAID creates an optimal storage foundation that fully leverages the capabilities of our enterprise NVMe technology.”
The implementation delivers significant operational advantages:
- Zero Single Points of Failure: Dual-port NVMe, mirrored metadata targets, and distributed object storage pools.
- Exceptional Power Efficiency: Half-rack footprint delivers top-tier performance with reduced power and cooling requirements.
- Commodity Economics: 100% off-the-shelf hardware eliminates vendor lock-in.
- Future-Ready Architecture: PCIe 5.0 and NDR 400 InfiniBand provide expansion headroom.
The Helma project represents one of the largest AI-focused storage deployments in European academia, with a total storage capacity of 4.9 petabytes across the 10 HA-pair-servers xiRAID cluster. The system supports concurrent access from 768 high-end GPUs processing multi-terabyte datasets, with workloads requiring sustained throughput of hundreds of gigabytes per second and millions of IOPS for metadata operations. This scale of deployment, achieving top-3 global IO500 performance in just half a rack, demonstrates the transformational potential of software-defined storage for large-scale AI infrastructure. The success of this implementation positions it as a reference architecture for next-generation academic and commercial AI computing facilities worldwide.
The complete case study, including detailed technical specifications and performance analysis, is available at https://xinnor.io/case-studies/helma/.
