Operations: Slurm, Kubernetes and fleet management

40 to 48 h · USD 550 with the exam

1. core AI Infrastructure and Operations Fundamentals Seven-hour overview of GPUs, DGX systems, networking, storage, power and cooling. Shared base for the associate and professional exams. free to audit7 h
2. core Slurm in containers with a GPU giovtorres/slurm-docker-cluster starts a full Slurm cluster with Docker Compose in minutes; its GPU image registers the card as a GRES automatically. The quickest way to a working cluster on one machine. free3 h
3. core Slurm administration: accounting, QoS, fairshare, topology SchedMD's own administrator guides, then practise with sacctmgr. University guides fill in worked fairshare examples.
   Also: QoS · Topology · Twente admin guide · HPC from scratch: accounting · SLUG talks free5 h
4. core Pyxis and Enroot NVIDIA's plugin that adds srun --container-image to Slurm. This is how GPU jobs run in containers on most neoclouds. free2 h
5. core GPU Operator, DCGM exporter, Prometheus and Grafana Install the GPU Operator on Kubernetes, export DCGM metrics, build a dashboard, then try time-slicing or MIG.
   Also: DCGM · GPU sharing free5 h
6. optional Kueue quick start Job queueing for Kubernetes, the closest thing to Slurm partitions in the Kubernetes world. free1.5 h
7. core Base Command Manager Administration course Self-paced course on NVIDIA's cluster provisioning tool. free8 h NVIDIA's Base Command Manager page currently says "Register for Free" for this self-paced course. An older NVIDIA DGX learning path still shows USD 50, so re-check before paying.
8. core Run Base Command Manager yourself NVIDIA licenses BCM free to organisations for their own use: up to 8 accelerators per node, any number of nodes, community support only. A head node and one or two compute VMs give real hands-on. Running it in VMs was not confirmed in the official docs. free4 h Opened 2026-10-06: the licensing page states the free licence covers up to eight accelerators per server with no limit on servers, is for an organisation's own use and not for resale, and comes without enterprise support. Running the head node and compute nodes as VMs was not confirmed in NVIDIA's documentation.
9. optional Run:ai: read, do not install No free or open self-hosted edition was found; only the scheduler core is open source as KAI Scheduler. Read the overview. free1 h
10. optional Slinky, Soperator, SUNK Three ways of running Slurm on Kubernetes (SchedMD, Nebius, CoreWeave). Know what they are; building one is not needed for the exam. free1 h
11. skip NVIDIA DeepOps Ansible bring-up for multi-node DGX clusters. Too heavy for a single-GPU lab. free
12. skip AI Operations Professional public training Instructor-led preparation at six times the exam price. USD 3,00024 h Opened 2026-10-06: NVIDIA's DGX data center learning path lists the AI Operations Professional instructor-led workshop at USD 3,000 for 24 hours: https://www.nvidia.com/en-us/learn/learning-path/dgx-data-center/
13. optional Exam: NCA-AIIO (associate) See the exam entry below. USD 1251 h
14. core Exam: NCP-AIO See the exam entry below. USD 5002 h