Hardware

Resources¶

Our cluster provides the following hardware resources.

CPUs¶

If you do not load any hardware modules, your job will end up on the first available node. The node may have either AMD or Intel CPU. If you need a specific CPU type, you would need to load the corresponding module, e.g. module load amd.

Nodes where the number of cores differs from the number of CPUs have simultaneous multithreading (SMT) enabled. This means multiple hardware threads (typically two) are available on each physical core. The cluster is configured to allocate resources at the physical core level but CPUs are allocated as hardware threads. Consequently, the number of allocated CPUs may be larger than requested. For example, a job requesting 5 CPUs will be allocated 6 CPUs on the nodes where SMT is enabled.

To request physical cores rather than CPUs, you can specify --hint=nomultithread. Slurm will schedule one hardware thread per core, allowing variables such as SLURM_CPUS_PER_TASK to correspond directly to the number of physical cores. Please note that billing is based on the number of allocated CPUs reported by Slurm. On nodes with SMT enabled, a job that requests 4 physical cores will be allocated 8 CPUs, and billing will be based on the 8 allocated CPUs.

For multi-node jobs that require fast internode communication, you may want to consider the Supercomputer service.

GPUs¶

There are different NVIDIA GPUs available: H200, H100, A100, L4. The H200, H100 and A100 nodes have nvlink (high performance bandwidth interconnect among GPUs). These machines are suitable for multi-GPU workloads. For multi-GPU jobs, you would need to request either A100, H100 or H200 specifically (e.g. module load h100). Alternatively, you can load the multigpu module, i.e. module load multigpu.

L4 GPUs are meant for single-GPU jobs only. If your job requires a lot of GPU memory, you may want to request a GPU flavor by VRAM, for example: module load gpumem80gb.

For GPU jobs with shorter durations, consider submitting with --partition lowprio.

For multi-GPU multi-node jobs that require fast internode communication, you may want to consider the Supercomputer service.

Hardware¶