DeltaAI

RP account needed

DeltaAI is NCSA's Grace Hopper supercomputer, designed for AI/ML training. It comprises 152 nodes built on the NVIDIA GH200 superchip.

Jobs

Jobs on DeltaAI are submitted through the Slurm scheduler. Users can run batch jobs with sbatch,
launch interactive work with srun or salloc, or access compute resources through Open OnDemand
applications such as JupyterLab.

Node sharing is the default behavior. The smallest allocatable unit is a single GH200 superchip - roughly 1 GPU, 72 CPU cores, and 187 GB of RAM - so jobs do not need to request a whole node. Users who need exclusive access can request all consumable resources for the node or use Slurm options such as --exclusive --mem=0. All nodes are of a single type and any job can span any node.

DeltaAI charges jobs based on the resources reserved for the job, not the resources used. One
Service Unit (GPU Hour) is one GH200 superchip for one hour, and the number of superchips charged is the largest of: the GPUs requested, the CPU cores requested divided by 72, or the memory requested divided by 187 GB. A request for 1 GPU with 300 GB of memory is therefore charged 2 SU per hour, and a whole node is charged 4 SU per hour. Note that the RAM available to jobs via the Slurm --mem option is 10 to 15 percent less than the total installed, because of memory reserved for the operating system.

If you do not specify them, jobs default to a 30-minute wall-clock limit and 1000 MB of memory per
core. Jobs are not automatically requeued or restarted by default; add --requeue to enable it.

For more information about running jobs on DeltaAI, see the NCSA DeltaAI [Running Jobs]
documentation.

Queue specifications

Metrics updated 2026-06-16

Name Purpose Nodes CPU cores / node GPUs / node Node RAM Jobs
30 days
Wait Time
30-day trend
Wall Time
30-day trend
ghx4 Default GPU queue with 4 GPUs per node for standard production jobs 152 NVIDIA Grace ARM (aarch64) (288 cores) 4 NVIDIA H100 (96 GB vRAM) 480 GB 136,931
ghx4 wait time: average 11.4 hours, range 2.1 to 41.2 hours over 30 days
ghx4 wall time: average 1.0 hours, range 0.2 to 3.2 hours over 30 days, wall-time limit 48h
ghx4-interactive Interactive GPU sessions for developing, debugging, and testing; max 4 nodes per job, 8 node-hours per user, charged at 2x 152 NVIDIA Grace ARM (aarch64) (288 cores) 4 NVIDIA H100 (96 GB vRAM) 480 GB 34,237
ghx4-interactive wait time: average 0.6 hours, range 0 to 1.9 hours over 30 days
ghx4-interactive wall time: average 0.1 hours, range 0 to 0.9 hours over 30 days, wall-time limit 2h

Software

No software usage data is currently reported for DeltaAI in XDMoD.


Storage

DeltaAI users have HOME, PROJECTS, and WORK file systems, plus per-node local /tmp. /work/hdd and /work/nvme are shared between Delta and DeltaAI. 

DeltaAI does not provide a /scratch file system.

Snapshots are only available for data in your /u directory and are retained for 30 days. These snapshots are not backups - you are responsible for backing up your files in /projects and /work.

File System

Directory Path Quota Purge Backup Notes
HOME /u 100 GB, 750,000 files Not purged Yes (30 days) Area for software, scripts, and job files. Not for I/O during jobs. Snapshots are retained 30 days and are not backups.
Projects /projects 500 GB, 300,000 files Not purged No Area for shared data for a project, common data sets, software, results, and so on. Up to 1-25 TB by allocation request.
Work - HDD /work/hdd 1 TB, 500,000 files Not purged No Area for computation, largest allocations, where I/O from jobs should occur. Best for large sequential I/O. Up to 1-100 TB by allocation request. Shared between Delta and DeltaAI.
WORK - NVMe /work/nvme 1 TB, 500,000 files Not purged No Area for computation, NVME is best for lots of small I/O from jobs. Up to 1-100 TB by allocation request. Shared between Delta and DeltaAI.
/tmp /tmp 3900 GB Purged after each job No Locally attached disk for fast small file I/O. Save results to other file systems before job ends.

External Storage

For large data transfers, Globus is recommended. For smaller transfers, SCP or Rsync are recommended.

 

Helpful Commands: The quota command is useful for viewing your usage of file systems quota. The accounts command is useful for mapping ACCESS projects to local projects on DeltaAI accounts


File Transfer

Users may use scp or rsync for small to modest transfers, while Globus is recommended for large data transfers.

Files can also be shared externally from the /projects file system using Globus. To share data,
users should create a sharing directory under their project space, move or copy the data into
that directory, and connect to the "ACCESS Delta" Globus endpoint to create the shared
collection.

For more information, see the NCSA DeltaAI [File Sharing] and [Transferring Data] documentation.

Supported Methods Data Transfer Node URL
GLOBUS | RECOMMENDED Globus
SCP | RECOMMENDED dtai-login.delta.ncsa.illinois.edu Transferring Files
RSYNC | RECOMMENDED dtai-login.delta.ncsa.illinois.edu Transferring Files