GPU-Cluster

• Term: Funded in 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 530648688

A total of 20 computing nodes, each with 4 GPGPU accelerators, are to be procured for the GPU cluster. Of these, 4 computing nodes are to be equipped with Nvidia H100 GPUs (80GB VRAM, PCIe) in order to enable calculations with a particularly large amount of RAM. The other 16 nodes are to be equipped with 4 Nvidia A40s each (or, if available, with the successor L40). A fast NVMe-based memory is also planned for I/O-heavy codes. In order to connect to the existing HPC cluster (in particular to the parallel GPFS storage system), it is planned to be equipped with HDR Infiniband. Since CPUs usually play a subordinate role in use, we aim to equip all nodes with one CPU socket, a CPU with at least 32 cores (64 threads) and at least 384 GB of main memory. Due to the very high heat generation to be expected, direct water cooling (DLC) is to be used for the entire system. This not only enables significantly better heat dissipation, which significantly reduces energy costs, but also better packing density, which minimizes space requirements.

• DFG Programme: Major Research Instrumentation
• Major Instrumentation: GPU-Cluster
• Instrumentation Group: 7040 Vektorrechner
• Applicant Institution: Universität Münster
• Leader: Professor Dr. Raimund Vogl