GPU nodes
GPU nodes
Section titled “GPU nodes”The UBELIX GPU-based hardware partition currently consists of about 20 compute nodes. Due to their different architectures, their specifications differ as shown in the table below.
| Nodes | GPUs | CPU cores | GPU Memory | Network |
|---|---|---|---|---|
| 2 | 8x NVIDIA RTX3090 | 32 (2x16) | 24 GB | 100Gb/s |
| 10 | 8x NVIDIA RTX4090 | 128 (2x64) | 24 GB | 100Gb/s |
| 1 | 8x NVIDIA RTX6000 | 128 (2x64) | 96 GB | 100Gb/s |
| 1 | 6x NVIDIA A100 | 128 (2x64) | 80 GB | 100Gb/s |
| 5 | 8x NVIDIA H100 | 128 (2x64) | 96 GB | 100Gb/s |
| 2 | 8x NVIDIA H200 | 128 (2x64) | 141 GB | 100Gb/s |
See the Slurm partitions page for an overview of options for allocating these nodes.
Overview
Section titled “Overview”Each GPU compute node is equipped with eight GPUs of a given type with up to 96GB GPU memory (H100) depending on the node type.
CPU & Memory
Section titled “CPU & Memory”Due to the different GPU compute node generations the amount of CPU core and memory per node differs. In terms of available memory per GPU this can be summarized as follows:
| GPU Type | CPUs per GPU | Memory per GPU |
|---|---|---|
| Nvidia RTX 3090 | 4 | 60GB |
| Nvidia RTX 4090 | 16 | 90GB |
| Nvidia A100 | 20 | 80GB |
| Nvidia H100 | 16 | 90GB |
| Nvidia H200 | 16 | 90GB |
Network
Section titled “Network”UBELIX GPU compute nodes each have a single 100 Gb/s interface to the Infiniband interconnect.
Disk storage
Section titled “Disk storage”Depending on the node type the CPU compute nodes are equipped with up to 1.5TiB of local NVMe SSD storage avaible as local scratch for high-performance I/O operations. See the UBELIX data storage options for details.