HubIA's DGX overview¶
Platform information¶
As of February 21, 2026:
- DGX OS: NVIDIA DGX Station A100 7.4.0
- Base OS: Ubuntu 24.04.4 LTS
- Kernel: Linux 6.8.0-100-generic
- GPUs: 4 × NVIDIA A100 (80 GB each) for compute + 1 display GPU
- CPU: AMD EPYC 7742, 64 physical cores / 128 logical cores
- System memory: 512 GB DDR4 RAM
- Data storage: 7.68 TB NVMe
- OS storage: 1.92 TB NVMe
GPU partitioning (MIG)¶
The DGX uses MIG (Multi-Instance GPU) to share GPUs across users and workloads.
Current pools are:
- 10 GB VRAM (standard GPU slice): 7 ×
1g.10gb - 40 GB VRAM (large GPU slice): 2 ×
3g.40gb - 80 GB VRAM (full GPU): 2 ×
A100 80GB
See GPU and MIG layout for the detailed mapping.
Access and scheduling model¶
- Users connect to the DGX through SSH.
- Compute resources are allocated by Slurm.
- Main commands are:
srunfor interactive sessionssbatchfor batch jobs
Available partitions are described in Slurm (quick guide). Technical partition policy is described in Advanced partitions. End-to-end user examples are in Use case example.