Cluster comparison¶

Side-by-side overview of the three DIPC HPC clusters to help you choose the right system for your workload.

Atlas FDR end of life

Atlas FDR operates with minimal maintenance and will be gradually decommissioned during 2026. New users should choose Hyperion or Atlas EDR.

Hardware overview¶

Feature	Hyperion	Atlas EDR	Atlas FDR
Compute nodes	302	85	225
Total cores	17,292	2,656	6,116
Total RAM	222 TB	16 TB	60 TB
GPUs	105 (A100, RTX 3090, RTX 4090, A6000, A30, RTX PRO 6000)	54 (P40, RTX 3090, RTX 8000, A30)	None
Interconnect	HDR InfiniBand (200 Gb/s)	EDR InfiniBand (100 Gb/s)	FDR InfiniBand (56 Gb/s)
Processor generations	Cascadelake, Icelake, Emerald Rapids, Zen 3	Skylake, Cascade Lake	Haswell, Broadwell, Skylake
Max cores per node	128	80	52
Max memory per node	4 TB	2 TB	1.5 TB

Feature	Hyperion	Atlas EDR	Atlas FDR
Home directory	/dipc (50 GB quota)	/dipc (no quota)	/dipc (no quota)
Scratch	/scratch (2 TB, BeeGFS)	/scratch (1.5 TB)	/scratch (500 GB)
Long-term storage	/data (10 TB, read-only on compute)	-	-
Local scratch	/lscratch (varies, up to 14 TB)	/lscratch (~750 GB)	/lscratch (~700 GB)
Home on compute nodes	Read-write	Symlink to /scratch	Symlink to /scratch
Backups	None	/dipc daily (2 months)	/dipc daily (2 months)

Feature	Hyperion	Atlas EDR	Atlas FDR
SLURM instance	Independent	Shared with Atlas FDR	Shared with Atlas EDR
MPI launcher	`srun`	`srun`	`mpirun`
Software stack	Latest toolchains	Shared modules	Older toolchains

I need GPUs for deep learning or CUDA workloads

Use Hyperion for A100 GPUs (best for large models and multi-GPU training) or Atlas EDR for RTX 3090 and P40 GPUs.

I need large memory per node (1 TB+)

Hyperion offers nodes with up to 4 TB of RAM.

I need to run large MPI jobs

Hyperion has the fastest interconnect (HDR, 200 Gb/s) and the most nodes, making it ideal for large-scale parallel jobs.

I am a new user and not sure where to start

Start with Hyperion. It has the most resources, the newest hardware, and supports all workload types.