Zwei GPU-Rechenknoten

null

The core scope of work and statement of work summary for this tender involve the procurement, delivery, and installation of two GPU compute nodes, including all necessary components and services, within a maximum budget of 1.2 million Euros (gross).

The key deliverables and responsibilities include:

• Provision of two identical GPU compute nodes, each equipped with 8 CUDA-compatible GPUs (at least 260 GB VRAM, 9 PetaFlops FP8 performance with sparse-tensor operations, GPU memory bandwidth >= 7.7TB/s), 1 or 2 CPUs with at least 64 x86_64 cores, and SSDs for the operating system (min. 480GB) and additional NVMe-SSDs for data.
• Integration of the modular system components exclusively into the provided 19-inch server rack (Knürr 42RU-Cabinet Miracel) at the RUB Data Center.
• Installation of the current version of Alma Linux 9.x as the operating system on all nodes.
• Installation and functional setup of a resource manager (SLURM) on all nodes, integrated with the existing Elysium Cluster.
• Pre-installation of a software stack including current GNU and LLVM compilers for C, C++, and Fortran, OpenMPI 5, and current CUDA and NVHPC versions.
• Inclusion of all necessary software licenses for system operation for at least five years.
• Network integration by connecting both nodes to existing Omnipath (Rack 02.10) and FS S5800-48T4S-PE Ethernet switches for service (>= 1 Gbit/s Ethernet) and management (>= 1 Gbit/s Ethernet, IPMI).
• Utilization of existing network services such as DNS, NTP, Slurm, and DHCP, and connection to LDAP for user and group management.
• Mounting of Elysium file systems (/home, /cluster, /lustre) on all nodes.
• Delivery to the point of use (RUB-DC), including transport, setup, complete commissioning, and immediate removal and disposal of packaging material.
• Execution of functional tests post-delivery as part of the acceptance procedure, including OSU, HPCG, and GPU-Burn benchmarks performed by the contractor on the installed system.
• Handover of all information and passwords required for operating and administering the cluster to the HPC@RUB Team upon system acceptance.
• The total power consumption of the complete cluster expansion must not exceed 40 kW under any load.