Zwei GPU-Rechenknoten

The documents do not explicitly state the contract start date, end date, total duration, or key milestones for the contract execution. However, the following dates related to the tender process are provided:

• Deadline for submitting bidder questions: 04.09.2026, 12:00 Uhr
• End of the offer period: 11.09.2026, 12:00 Uhr
• Currently planned award of contract by: 16.10.2026
• Binding period for offers until: 16.11.2026

The core scope of work for this tender involves the acquisition, delivery, installation, and full commissioning of two GPU computing nodes. These nodes are intended to expand an existing High-Performance Computing (HPC) cluster at the Ruhr-Universität Bochum to facilitate working with larger language models (LLMs).

Key deliverables and responsibilities include:

• Supply of Hardware: Two identically equipped GPU computing nodes, each featuring 8 CUDA-compatible GPUs with at least 260 GB VRAM, specific performance metrics (9 PetaFlops FP8 Performance for sparse-Tensor operations, GPU memory bandwidth >= 7.7TB/s), 1 or 2 CPUs with at least 64 x86_64 cores, an SSD for the operating system (min. 480GB), and an additional NVMe-SSD for data.
• Integration and Compatibility: The components must be integrated into a provided 19"-server rack (Knürr 42RU-Cabinet Miracel) within the RUB Data Center, ensuring compatibility and that the total power consumption does not exceed 40 kW. This also includes integration with existing Lustre Client and Slurm software, existing network services (DNS, NTP, Slurm, DHCP), and LDAP for user/group management. Elysium file systems (/home, /cluster, /lustre) must be integrated.
• Software Installation: Installation of Alma Linux 9.x, a functional resource manager (SLURM) integrated with the Elysium Cluster, and a software stack including current GNU and LLVM compilers (C, C++, Fortran), OpenMPI 5, and current CUDA and NVHPC versions. All necessary software licenses for five years of operation must be included.
• Delivery and Installation Services: Transport, setup, and full commissioning of the systems at the RUB-DC premises. This also covers the immediate removal and professional disposal of packaging material, and the labeling of all network and power cables.
• Networking: Connection of the nodes to an existing Omnipath Switch via high-performance optical cables and transceivers, as well as connections to existing service and management network switches with appropriate cabling.
• Functional Testing: Performance of required benchmarks (Osu, HPCG, GPU-Burn) by the contractor as part of the acceptance test to ensure specified bandwidth and system stability through 24-hour burn-in tests.
• Documentation and Information Transfer: Provision of a system description, component lists, power consumption details, network and setup diagrams, and a script for BMC interface login credentials. All operational and administrative information, including passwords, must be handed over to the HPC@RUB Team upon acceptance.
• Budget Constraint: The maximal budget of 1.2 million Euro gross must not be exceeded.