GPU and MIG layout

This page describes the current DGX GPU partitioning used by Slurm.

Current layout

As of February 21, 2026:

Resource pool Quantity Used by partition(s)
A100 80GB full GPUs (nvidia_a100-sxm4-80gb) 2 prod80
3g.40gb MIG instances (nvidia_a100_3g.40gb) 2 prod40
1g.10gb MIG instances (nvidia_a100_1g.10gb) 7 interactive10, prod10
DGX display GPU 1 system/display (not a user compute partition)

Relation with Slurm partitions

  • interactive10: one 1g.10gb MIG per job (interactive usage).
  • prod10: one 1g.10gb MIG per batch job.
  • prod40: one 3g.40gb MIG per batch job.
  • prod80: one full A100 80GB GPU per batch job.

For submission syntax and user workflow, see Slurm (quick guide). For technical policy (defaults/caps), see Advanced partitions.

How to verify on the node

nvidia-smi -L
scontrol show node $(hostname -s) | egrep 'Gres=|CfgTRES|State='

If GPU partitioning changes, this page and the partition pages must be updated together.