NVIDIA's NCP-AIO study guide splits the exam into two sections and gives each half of the score. One is multiple choice. The other is a performance-based section of three challenges, each a short scenario you solve in a live virtual machine. The study guide names what those challenges expect you to handle: using Slurm in Base Command Manager (BCM) to troubleshoot a cluster, using Kubernetes to run a workload, and using BCM to add a node to a cluster. As of October 2026, NVIDIA's certification page describes the exam as 30 multiple-choice questions and 3 hands-on lab exercises in 120 minutes.
A question bank on its own prepares you for half of that score, so NCP-AIO preparation also needs a cluster you can break and repair.
The multiple-choice half
The written questions range across the four areas of the study guide. Administration covers Fleet Command, Slurm, Run:ai, MIG and data center design for AI. Workload management covers Kubernetes, BCM provisioning and the system tools DCGM, NVSM and nvidia-smi. Installation covers BCM itself, containers from NGC, cloud VMI containers, storage planning and DOCA services on BlueField DPUs. Troubleshooting covers Docker, the fabric manager for NVLink and NVSwitch systems, BCM, Magnum IO and storage performance.
One of the four areas is troubleshooting, and the other three are written as tasks, so the most useful practice questions describe a broken or half-configured system and ask for the next step.
Three NCP-AIO-style questions
On an 8-GPU server with NVSwitch, nvidia-smi lists all eight GPUs, but every CUDA job fails at startup with a "system not yet initialized" error. The NVIDIA driver was upgraded last night. What should the administrator check first?
On systems where GPUs connect through NVSwitch, Fabric Manager sets up the NVSwitch fabric, and CUDA cannot start work on the GPUs until it is running. Its package version has to match the driver version, so upgrading the driver without upgrading Fabric Manager leaves the service unable to start. A driver upgrade does not enable MIG. The DCGM exporter publishes metrics and plays no part in starting CUDA jobs. A container runtime setting would affect containers only, while here every CUDA job fails.
A GPU in node gpu07 failed a health check, and Slurm set the node to DRAIN. The GPU has been replaced and passes diagnostics. Which command returns the node to service?
A drained node stays drained until an administrator clears the state, and scontrol update with State=RESUME does that. Restarting slurmd makes the node register again, but the drain recorded by the controller remains. sinfo reports node states and has no resume option. scancel cancels jobs and leaves the node state alone.
An administrator wants to split an A100 80GB into isolated GPU instances for several small inference services. Which step has to come first?
MIG partitions exist only after MIG mode is enabled on the GPU, and enabling it requires that no processes are using the GPU and that the GPU is reset. After that you create GPU instances and then compute instances inside them, so option C has the order reversed. MPS lets processes share a GPU without hardware isolation, which the requirement rules out. MIG on bare metal needs no vGPU license.
Full NCP-AIO set: 7 timed exams, every answer explained, $19.99 one-time.
See the practice testsOn the clock, the multiple-choice section shares 120 minutes with the three challenges, and NVIDIA's pages give no split between them. Our suggestion: answer the 30 questions at about 90 seconds each, which takes 45 minutes and leaves 75 for the labs. To practice that pace on a 60-question test, set a 90-minute limit.
Try 15 NCP-AIO practice questions free
They come from the NCP-AIO practice tests, and each option shows its explanation once you answer.
- 7 timed practice exams
- Every answer explained
- A score for each domain
- $19.99 one-time
The hands-on half
Troubleshooting a cluster with Slurm in BCM
Slurm's commands work the same on any cluster, and Slurm is open source, so two small virtual machines running a controller and a compute node are enough to practice reading sinfo and squeue, inspecting nodes with scontrol show node, and draining and resuming nodes. The BCM side, where Slurm is configured and managed through BCM, needs BCM itself. The study guide maps these objectives to units of NVIDIA's AI Operations course on provisioning with BCM and scheduling with Slurm, and it also lists a self-paced Base Command Manager Administration course. For practice outside a course, NVIDIA DSX Air (formerly NVIDIA Air) lists Base Command Manager among the software it can simulate, with a free trial as of October 2026.
Running a workload on Kubernetes
Kubernetes runs anywhere, including a laptop, which makes this the easiest challenge to rehearse. preporato.com Pro has Kubernetes labs mapped to NCP-AIO, among them installing the NVIDIA GPU Operator on k3s on a cloud GPU, diagnosing a GPU pod that stays Pending, resource requests and limits, priority and preemption and rolling, canary and blue-green updates. Each one is checked step by step, so a mistake shows up while you are still in the terminal.
Adding a node with BCM
Node provisioning happens in BCM, so it calls for the same environments as the BCM side of the Slurm challenge: NVIDIA's courses or a DSX Air simulation. The study guide's reading list for this area points to NVIDIA's Base Command Manager documentation and the DGX SuperPOD guide to provisioning nodes.
Where each option fits
| Option | Multiple-choice half | Hands-on half |
|---|---|---|
| NVIDIA's NCP-AIO study guide (free, linked from the certification page) | The objectives and a reading list for every area | Names the three challenge types |
| NVIDIA's AI Operations Professional Workshop | Listed on the certification page as recommended training | The study guide cites AI Operations course units on BCM provisioning and Slurm scheduling |
| NVIDIA DSX Air (sign-up) | Simulates BCM, Run:ai and NVIDIA networking; free trial as of October 2026 | |
| preporato.com NCP-AIO practice tests | Seven timed tests of 60 questions, an explanation for every option, a score for each exam area | Not covered: multiple choice only |
| preporato.com Pro labs | Kubernetes workloads, plus GPU operations labs on cloud GPUs; no BCM or Slurm | |
| Your own Slurm cluster on two virtual machines | Slurm commands for the troubleshooting challenge |
We make the preporato.com tests and labs, so treat our own rows with that in mind. The free questions on the certificate page come from the same tests, so you can judge the question style before paying.
If time is short, give the hands-on challenges at least as many hours as the question bank, since they carry the same weight in the score.
The 6-week NCP-AIO study plan orders the areas week by week, and the NCP-AIO cheat sheet summarizes the concepts and commands.
Frequently asked questions
Does NVIDIA offer an official NCP-AIO practice exam?
As of October 2026, NVIDIA's certification page lists the study guide, the AI Infrastructure and Operations Fundamentals course and the AI Operations Professional Workshop as preparation, and no practice exam.
Is the NCP-AIO exam still multiple choice only?
No. As of October 2026 it has 30 multiple-choice questions and 3 hands-on lab exercises, and the study guide gives each section half of the score. Guides and question banks that describe NCP-AIO as multiple choice only were written for the earlier format.
Do I need my own GPUs to practice the hands-on challenges?
Slurm and basic Kubernetes practice run on ordinary virtual machines without GPUs. Anything that touches the GPU itself, such as the GPU Operator, MIG or DCGM, needs one, and the preporato.com Pro labs run those steps on cloud GPUs.
How do the preporato.com NCP-AIO tests compare with the real multiple-choice section?
Each of our tests has 60 questions and allows 120 minutes, twice the length of the real section, and draws on the same four areas of the study guide. Taking one at the 90-second pace above is a good rehearsal for the written half.