Skip to content

Available Resources

Available Resources

Name Partition CPU specs GPU specs RAM Other notes
node-accel-1 accel AMD EPYC 9555 @ 3.0 GHz, 128 CPUs 8 x NVIDIA H200 NVL 3094.00 GB scratch
node-epyc-1 power AMD EPYC 9755 @ 2.70 GHz, 256 CPUs - 750.00 GB scratch
node-epyc-2 AMD EPYC 9755 @ 2.70 GHz, 256 CPUs - 750.00 GB scratch
node-epyc-3 AMD EPYC 9755 @ 2.70 GHz, 256 CPUs - 750.00 GB scratch
node-epyc-4 power, power-long AMD EPYC 9755 @ 2.70 GHz, 256 CPUs - 750.00 GB scratch
node1 weakold Intel Xeon CPU E-2650 0 @ 2.00 GHz, 16 CPUs - 61.44 GB
node2 Intel Xeon CPU E-2650 0 @ 2.00 GHz, 16 CPUs - 61.44 GB
node4 Intel Xeon CPU E-2650 0 @ 2.00 GHz, 16 CPUs - 61.44 GB
node5 Intel Xeon CPU E-2650 0 @ 2.00 GHz, 16 CPUs - 125.95 GB
node6 Intel Xeon CPU E-2650 0 @ 2.00 GHz, 16 CPUs - 125.95 GB
node7 Intel Xeon CPU E-2650 0 @ 2.00 GHz, 16 CPUs - 253.95 GB
node8 Intel Xeon CPU E-2650 0 @ 2.00 GHz, 16 CPUs - 253.95 GB
node10 regular Intel Xeon Gold 5220R CPU @ 2.20 GHz, 48 CPUs - 1156.34 GB
node11 Intel Xeon Gold 5220R CPU @ 2.20 GHz, 48 CPUs - 378.88 GB
virtnode1 debug Intel Xeon Gold 5318Y CPU @ 2.10 GHz, 6 CPUs - 11.26 GB 2:00:00 Timelimit

Important information

High-performance storage (scratch)

Scratch storage

Scratch storage is intended to speed up computations when large amounts of data need to be read and written.

Key points to keep in mind about scratch disks
  1. Available only on specific nodes (in Slurm marked with Feature=scratch).
  2. Similar to /home, each user has their own personal scratch directory under /scratch.
  3. Scratch disk contents are not synchronized automatically; users must copy data at the start and end of the job.
  4. Files that have not been used for a week are automatically deleted from scratch disks.

Submitting GPU jobs

If your job requires GPU resources, you must specify the number of GPUs requested in your SBATCH script (and optionally the GPU type).

Currently available GPU resource types
  • h200
  • h200_1g.18gb
Parameters

-G, --gpus=[type:]<number>

Example
1 GPU (any type):          --gpus=1
1 H200 GPU:                --gpus=h200:1
2 H200 GPUs:               --gpus=h200:2
1 H200 18GB MIG instance:  --gpus=h200_1g.18gb:1
2 H200 18GB MIG instances: --gpus=h200_1g.18gb:2
GPU partitioning with NVIDIA MIG

2 of the 8 H200 GPUs on node-accel-1 are divided into smaller, mutually isolated GPU instances using NVIDIA MIG (Multi-Instance GPU) technology.

MIG allows a single physical GPU to be divided into several independent instances, each with its own GPU resources and video memory. This allows multiple jobs to use the same physical H200 GPU simultaneously when each job does not require the full GPU performance or memory capacity.

node-accel-1 GPU layout:

  • 6 full H200 GPUs
  • 14 H200 MIG instances with approximately 18 GB of memory each

Partitions

Jobs can be submitted to five different partitions by setting the partition parameter in Slurm scripts: --partition=<name>

The debug partition is intended for short test jobs. It is designed to be readily available.

  • Default runtime: 2 hours
  • Maximum runtime: 2 hours

The regular partition is intended for typical compute jobs.

  • Default runtime: 24 hours
  • Maximum runtime: 240 hours

The power partition is intended for high-performance CPU compute jobs.

  • Default runtime: 2 hours
  • Maximum runtime: 72 hours

High-performance storage is available on this partition.

The power-long partition is intended for long-running high-performance CPU compute jobs.

  • Default runtime: 24 hours
  • Maximum runtime: 240 hours

High-performance storage is available on this partition.

The accel partition is intended for GPU-accelerated compute jobs.

  • Default runtime: 2 hours
  • Maximum runtime: 48 hours

GPU jobs must specify the GPU count (see Submitting GPU jobs).

High-performance storage is available on this partition.

The weakold partition consists of relatively old and lower-performance compute servers, intended for testing or low-priority jobs.

  • Default runtime: 24 hours
  • Maximum runtime: 240 hours

Archiving

Less frequently used files can be archived by moving them from /home/<username> to /archive/<username>.