NVwulf has three kinds of GPU node, and each has a standard queue, a -long queue and, for two of them, a debug- queue for quick tests. Every NVwulf node is shared: several people's jobs can run on the same node, so ask for exactly the GPUs, CPUs and memory you need.

On this page: Queues and limits · Default CPUs and memory · Priority queues

Queues and limits

QueueGPUGPUs per nodeCPUs per nodeHost memoryLongest runMax nodes
debug-h200x4H200 NVL, 141 GB464 (62 usable)about 750 GB1 hour1
debug-b40x4RTX PRO 6000 Blackwell, 96 GB464 (62 usable)about 512 GB1 hour1
h200x4H200 NVL464 (62 usable)about 750 GB8 hours2
h200x4-longH200 NVL464 (62 usable)about 750 GB48 hours1
h200x8H200 NVL864 (62 usable)about 1,500 GB8 hours1
h200x8-longH200 NVL864 (62 usable)about 1,500 GB48 hours1
b40x4RTX PRO 6000 Blackwell464 (62 usable)about 512 GB8 hours2
b40x4-longRTX PRO 6000 Blackwell464 (62 usable)about 512 GB48 hours1

Start with a debug- queue for short tests. Open OnDemand apps on NVwulf run in these same queues; see Open OnDemand apps on NVwulf.

Default CPUs and memory

If you don't ask for CPUs or memory, Slurm gives your job a default based on the number of GPUs you request. To see the defaults and limits for a queue, run:

scontrol show partition h200x4

To list every queue you can use with its time limit:

sinfo -o "%P %l %D"

Priority queues

Users and groups with a dedicated allocation on NVwulf also have priority queues. They run on the same kinds of hardware with the same limits as the standard queue they mirror, but their jobs are scheduled ahead of general jobs:

  • p-h200x4 and p-h200x4-long
  • p-h200x8 and p-h200x8-long
  • p-b40x4 and p-b40x4-long
  • p-h200x8-03-long: 8 H200 SXM GPUs, 48 hours

To ask about a dedicated allocation, contact HPC support.

Related: Getting started on NVwulf · Fairshare and job priority · How long can a job run? · NVwulf

Applies to NVwulf