8 of the 15 questions in this set, with the correct answer marked and every option explained.
1. Which NVIDIA product provides cluster provisioning, node image management, user and role administration, and a Base View interface for real-time monitoring of an AI cluster?
✓NVIDIA Base Command Manager (BCM)
BCM is the cluster management platform at the centre of this exam. Provisioning and imaging nodes, organizing them into categories, managing users and permissions, configuring networking, and monitoring through Base View are all BCM responsibilities named directly in the blueprint.
✗NVIDIA Triton Inference Server
Triton serves models for inference. It runs as a workload *on* the cluster; it does not provision or administer the cluster.
✗NVIDIA NeMo
NeMo is the framework for building, customizing and deploying generative AI models. Again, a workload rather than the management layer.
✗NVIDIA Nsight Systems
Nsight Systems profiles an application's performance timeline. It is a developer tool with no cluster administration function.
2. In Base Command Manager, what is the purpose of organizing compute nodes into categories?
✓To apply a common configuration, software image and settings to a group of nodes that share hardware characteristics or workload role.
Categories are BCM's mechanism for managing nodes at scale. Rather than configuring each node individually, nodes with the same hardware or purpose inherit a category's image and settings, so changes propagate consistently.
✗To determine physical rack placement during installation.
Categories are a logical configuration grouping. Physical placement is documented separately and is not what BCM categories express.
✗To set the billing rate charged to each research team.
Chargeback and accounting come from the workload manager's job records, not from node categories.
✗To define the network topology between switches.
Switch topology is a networking design concern. BCM configures node network settings, but categories group nodes rather than describing fabric topology.
3. A team is deploying DOCA Services on a BlueField DPU. Where do those services actually run?
✓On the Arm cores of the DPU itself, isolated from the host operating system.
A BlueField DPU carries its own Arm CPU complex running its own operating system. DOCA Services are deployed there, which is what gives the DPU an infrastructure control plane isolated from — and not consuming cycles of — the host.
✗On the host x86 CPU, with the DPU acting only as a network interface.
This would defeat the purpose. Offloading infrastructure services *off* the host CPU is the DPU's central value proposition.
✗On the GPU's streaming multiprocessors.
GPUs execute compute kernels for AI and HPC workloads. They do not host infrastructure services or a general-purpose operating system.
✗On the top-of-rack switch's control processor.
Switch processors run switch firmware such as Cumulus Linux. DOCA Services are a DPU deployment target, not a switch one.
4. Which statement best describes what NVIDIA Run:ai adds to a Kubernetes-based AI cluster?
✓GPU-aware scheduling, quota and fair-share management, and workload orchestration that lets teams share GPU capacity efficiently.
Kubernetes on its own treats GPUs as a fairly coarse resource. Run:ai layers on GPU-aware scheduling, guaranteed and over-quota allocations per team, and fractional or pooled GPU usage — which is why the blueprint lists installing, administering and allocating resources with it.
✗A replacement container runtime that supersedes containerd.
Run:ai works with the existing Kubernetes runtime. It schedules and governs workloads rather than replacing the layer that executes containers.
✗A monitoring dashboard that replaces DCGM for GPU telemetry.
Run:ai surfaces utilization information, but DCGM remains the GPU health and telemetry layer. Scheduling and quota are Run:ai's purpose.
✗A model registry for storing trained checkpoints.
Artefact storage and model registry are separate concerns. Run:ai governs compute allocation, not model versioning.
5. An administrator must let four independent inference workloads share one H100 GPU with hardware-enforced isolation of memory and compute. Which configuration achieves this?
✓Configure Multi-Instance GPU (MIG) to partition the GPU into separate instances.
MIG creates hardware-isolated partitions, each with its own SM slice, memory and cache paths. That satisfies the requirement for enforced isolation between tenants sharing a single physical GPU, and MIG configuration is named directly in the Administration domain.
✗Run all four workloads in the default time-sliced sharing mode.
Time-slicing lets workloads share a GPU but provides no memory or fault isolation — one workload can exhaust memory and disrupt the others. The requirement specifically asks for hardware isolation.
✗Enable NVLink between the GPU and its neighbour.
NVLink joins GPUs into a larger pooled resource. The requirement is to subdivide one GPU, which is the opposite operation.
✗Assign each workload its own CUDA stream.
Streams provide concurrency within a process, not isolation between tenants. There is no enforced boundary between streams.
6. In a Slurm cluster, which command submits a batch job script to the scheduler queue?
✓sbatch
`sbatch` submits a batch script for later execution by the scheduler and returns immediately with a job ID. It is the standard submission path for non-interactive training jobs.
✗srun
`srun` launches a job step, typically running interactively in the foreground or launching tasks within an allocation. It is used inside batch scripts and for interactive work, not for queuing a script.
✗squeue
`squeue` displays the state of jobs already in the queue. It reports; it does not submit.
✗scancel
`scancel` terminates a queued or running job. It is the opposite operation.
7. What does the NVIDIA container toolkit provide?
✓The integration that lets containers access host GPUs, exposing the driver and devices into the container runtime.
Without it, a container has no path to the GPU. The toolkit hooks into the container runtime to inject the required device nodes and driver libraries, which is what makes `--gpus all` work.
✗A registry of pre-built AI container images.
That is NGC, NVIDIA's catalog. The toolkit is the runtime integration, not the image source.
✗A build system that compiles CUDA code inside containers.
Compilation is the CUDA toolkit's job. The container toolkit is about runtime GPU access.
✗A Kubernetes scheduler plugin for GPU bin-packing.
GPU scheduling in Kubernetes comes from the device plugin and schedulers such as Run:ai. The container toolkit operates one layer below, at the container runtime.
8. A team wants to pull a pre-built, GPU-optimized PyTorch container rather than building their own image. Where should they get it?
✓The NGC catalog
NGC is NVIDIA's catalog and registry of GPU-optimized containers, pretrained models, Helm charts and SDKs. Pulling framework containers from NGC is named directly in the blueprint under workload management.
✗The Base Command Manager node image repository
BCM images are operating-system images for provisioning bare-metal nodes. They are not application containers.
✗The DCGM metrics endpoint
DCGM exposes GPU telemetry. It distributes no software.
✗The UFM fabric database
Unified Fabric Manager manages InfiniBand fabrics. It has no container distribution role.
7 more questions in the app
Practise the full 15-question set with a timer, scoring and progress tracking.