8 of the 15 questions in this set, with the correct answer marked and every option explained.
1. After bringing up a new GPU cluster, an engineer needs to verify east-west fabric bandwidth between GPUs across nodes. Which test is designed for that?
✓An NCCL bandwidth test such as all_reduce_perf across multiple nodes.
NCCL implements the collective operations that distributed training actually uses, so an NCCL benchmark measures the real path — GPU to GPU, across NVLink inside a node and across the fabric between nodes. The blueprint names running NCCL to verify east-west fabric bandwidth explicitly.
✗High-Performance Linpack (HPL).
HPL is a dense linear algebra benchmark that stresses floating-point compute and is the standard test for raw FLOPS and thermal stability. It is used on this exam too, but for compute validation rather than fabric bandwidth.
✗A single-node stress test.
By definition a single-node test never crosses the fabric, so it cannot measure east-west bandwidth between nodes.
✗A storage throughput benchmark.
Storage tests measure the north-south path to the filesystem. East-west refers to node-to-node traffic between peers.
2. What is the purpose of running a burn-in test such as an extended HPL or NCCL run during cluster acceptance?
✓To hold the cluster under sustained load long enough to surface intermittent faults — thermal throttling, marginal cables, failing components — that a short test would miss.
Burn-in is about duration. Many real faults only appear once components reach steady-state temperature or after enough operating hours, so a sustained run catches problems before production workloads do, while the hardware is still under acceptance.
✗To calibrate the GPU clock frequencies for the workloads that will run later.
Clock behaviour is managed by the driver and firmware according to thermal and power headroom. Burn-in does not calibrate it.
✗To pre-populate the filesystem cache so later jobs start faster.
Cache warming is transient and irrelevant to acceptance testing.
✗To satisfy a software licensing requirement before production use.
Burn-in is an engineering validation practice, not a licensing gate.
3. Which tool does NVIDIA's blueprint name for performing a multifaceted assessment of a cluster's nodes, covering aspects such as bandwidth, latency and node-to-node consistency?
✓ClusterKit
ClusterKit is NVIDIA's multi-test cluster assessment tool, running a suite of checks across nodes to surface outliers and inconsistencies. The blueprint's phrasing — 'run ClusterKit to perform a multifaceted node assessment' — matches this description directly.
✗ibdiagnet
ibdiagnet is an InfiniBand fabric diagnostic. It is genuinely useful here and appears on the NCP-AIN blueprint, but it examines the fabric rather than performing a broad multi-aspect node assessment.
✗nvidia-smi
nvidia-smi queries and manages individual GPUs on the local host. It is essential but single-node and single-purpose.
✗Helm
Helm packages and deploys Kubernetes applications. It performs no hardware assessment.
4. During cabling validation, an engineer needs to confirm not merely that a link is up but that its signal quality is adequate. Which measurement is most relevant?
✓Link-level signal quality indicators such as bit error rate and effective errors reported per port.
A marginal cable or transceiver commonly links up at full speed while quietly accumulating errors, which then degrades collective performance under load. Checking error counters and signal quality — rather than just link state — is what distinguishes a proper validation from a cursory one.
✗The physical length of the cable measured with a tape.
Length matters when specifying cables, but a within-spec cable can still be damaged or badly seated. Length tells you nothing about the signal actually carried.
✗Whether the cable's outer jacket colour matches the site standard.
Colour coding aids operations and documentation but carries no signal information.
✗The port's assigned VLAN ID.
VLAN configuration is logical network segmentation, entirely separate from physical layer signal integrity.
5. What does out-of-band management via a BMC provide during server bring-up?
✓Access to power control, console, sensors and firmware management independent of the host operating system, including while the host is powered off or unresponsive.
That independence is the entire point. A baseboard management controller runs on standby power with its own network path, so an engineer can power-cycle, view console output, read sensors and update firmware on a machine that has no working OS — essential during bring-up when there may be no OS yet.
✗Faster data transfer for training workloads by bypassing the host network stack.
That describes RDMA on the data network. The BMC's management interface is deliberately low-bandwidth and carries no workload traffic.
✗Direct GPU memory access for debugging CUDA applications.
Application debugging is done with developer tools such as cuda-gdb and Nsight. The BMC has no visibility into GPU memory.
✗Automatic scheduling of jobs when nodes become idle.
Job scheduling is the workload manager's role. The BMC has no awareness of jobs.
6. In the context of AI factory networking, what does a rail-optimized topology mean?
✓Each GPU position within every server connects to its own dedicated switch plane (rail), so corresponding GPUs across nodes communicate over a consistent, low-contention path.
Rail-optimized designs align the network to the collective communication pattern: GPU 0 on every node reaches GPU 0 on every other node through the same rail. That keeps all-reduce traffic predictable and reduces contention compared with an arbitrary mapping.
✗Servers are mounted on sliding rails to simplify physical maintenance.
This is the mechanical meaning of 'rail' and a deliberate distractor. In fabric design a rail is a switch plane, not a mounting bracket.
✗All network traffic is forced through a single aggregated link to simplify monitoring.
That would create a severe bottleneck. Rail-optimized topologies increase parallel paths, not reduce them.
✗The storage network and compute network share the same physical cabling.
Converging storage and compute traffic is a separate design decision, unrelated to what rail optimization means.
7. Before installing GPUs into a new rack of servers, which validation should be completed first?
✓Confirm that power delivery capacity and cooling capability meet the fully populated rack's requirements.
Power and cooling are prerequisites, not afterthoughts. Populating a rack that exceeds its circuit capacity or its cooling envelope leads to tripped breakers or sustained thermal throttling, and the blueprint lists validating power and cooling parameters as a bring-up step.
✗Install the deep learning frameworks the team intends to use.
Software installation comes much later, after hardware is validated and the control plane is in place.
✗Create user accounts for the research teams.
Account administration belongs to the operations phase once the cluster runs.
✗Submit a test job to the scheduler.
There is no scheduler yet at this point in bring-up, and no validated hardware for it to schedule onto.
8. Which command-line utility is the standard first stop for querying GPU presence, driver version, temperature, power draw and running processes on a single host?
✓nvidia-smi
The NVIDIA System Management Interface is the universal first check on any GPU host. It confirms the driver is loaded, enumerates GPUs, and reports utilization, memory, temperature, power and process occupancy.
✗ibstat
ibstat reports InfiniBand host channel adapter state — port status, rate and link layer. Useful on this hardware, but it says nothing about GPUs.
✗lsblk
lsblk lists block storage devices. It has no GPU awareness.
✗dmidecode
dmidecode reads DMI/SMBIOS tables for motherboard, memory and BIOS details. It will not report GPU driver state or utilization.
7 more questions in the app
Practise the full 15-question set with a timer, scoring and progress tracking.