42 free NVIDIA NCP-AII practice questions from the CertStash pack of 170, each with the correct answer and a full explanation. The rule is the same for every exam: ten free on the product page, 42 free here, all 170 in the pack, and each step includes the one before it.
Get all 170 questions (US$39) · Download the first 10 as a PDF
Question 1
A system engineer needs to set the vGPU scheduling behavior for all GPUs to share the scheduling equally with the default time slice length.
What command should be used?
Show answer and explanation
Correct answer: B. esxcli system module parameters set -m nvidia -p “NVreg_RegistryDwords=RmPVMRL=0x01”
On VMware vSphere, the vGPU scheduling policy for all GPUs is set through the RmPVMRL registry key, passed to the nvidia kernel module with esxcli system module parameters set -m nvidia -p followed by the NVreg_RegistryDwords string. A value of 0x01 selects the equal share scheduler with the default time slice length, so every running vGPU receives an equal share of GPU processing time. A value of 0x00 selects the best effort scheduler, which is the default behavior. The new setting takes effect after the ESXi host is rebooted.
Why the other options are wrong
- A. RmPVMRL=0x00 selects the best effort scheduler, which does not divide GPU time equally among vGPUs.
- C. The esxcli graphics namespace does not set kernel module parameters, which are configured under esxcli system module parameters.
- D. FRL refers to the frame rate limiter and is not the registry key that selects the vGPU scheduling policy.
Question 2
During a multi-day NeMo burn-in, intermittent “GPU fell off bus” errors occur.
Which diagnostic approach isolates hardware faults?
Show answer and explanation
Correct answer: A. Run DCGM diagnostics alongside burn-in to monitor GPU health metrics
A GPU that falls off the bus is a hardware or platform level fault, so the right approach is to collect GPU health data while the stress is applied. Running DCGM diagnostics and health monitoring alongside the NeMo burn-in records Xid errors, PCIe and NVLink errors, memory errors, temperatures and power behavior for each GPU. Correlating these readings with the moment of each failure shows whether a specific GPU, slot or board is at fault, which allows the affected component to be isolated and replaced instead of masking the symptom.
Why the other options are wrong
- B. Changing the model being trained alters the workload but provides no data about which hardware component is failing.
- C. HPL_USE_NVSHMEM is an HPL benchmark setting and has no role in diagnosing GPU faults during a NeMo burn-in.
- D. Lowering memory pressure may hide the symptom but does not identify the failing GPU or component.
Question 3
An engineer needs to verify the current firmware versions of all components (ATF, BSP, NIC, UEFI) on a BlueField-3 DPU’s BMC.
Which Redfish API command provides this information?
Show answer and explanation
Correct answer: C. curl -k -u root:<password> -X GET https://<bmc_ip>/redfish/v1/UpdateService/FirmwareInventory
The BlueField-3 BMC exposes the versions of the firmware components it manages through the standard Redfish UpdateService. A GET request to /redfish/v1/UpdateService/FirmwareInventory returns a collection with one member per component, such as the ATF, BSP (BlueField Board Support Package), NIC firmware, UEFI and the BMC itself, and each member reports its current version. Querying this endpoint with curl and the BMC credentials gives a single, authoritative view of all installed firmware versions, which is the standard check before and after a firmware update on the DPU.
Why the other options are wrong
- A. mstflint queries the firmware of a single NIC device and is not a Redfish call to the BMC.
- B. FirmwareList is not a resource defined under the Redfish UpdateService, so this request does not return the inventory.
- D. mlxconfig queries device configuration parameters, not the firmware versions of the DPU components managed by the BMC.
Question 4
An engineer needs to verify NVLink isolation on a single node with 8 GPUs.
Which NCCL test configuration stresses switch bisection bandwidth?
Show answer and explanation
Correct answer: B. Use all_reduce_perf -b 8 -e 16G -f2 -g 8 with NCCL_TESTS_SPLIT=“AND 0x1”
The nccl-tests suite can partition ranks into smaller groups that run the same operation in parallel through the NCCL_TESTS_SPLIT variable. With NCCL_TESTS_SPLIT set to AND 0x1, the ranks are divided into two groups, each using every other rank, and both all-reduce operations run at the same time. Because the two groups contend for the NVSwitch fabric simultaneously, the test exercises the bisection bandwidth of the switch and reveals whether each subset of GPUs is properly isolated and still reaches its expected bandwidth. The message size sweep from 8 bytes to 16 GB with -f 2 covers both latency and bandwidth regimes.
Why the other options are wrong
- A. DIV 8 places all eight GPUs of the node in one group, so no parallel partitioned operations are created to load the switch bisection.
- C. Without a split, a single all-reduce runs across all eight GPUs and does not create the competing groups needed to test bisection.
- D. Running on only four GPUs leaves half the node idle and cannot exercise bisection across the full eight GPU fabric.
Question 5
You are validating the environment of an NVIDIA GPU-accelerated data center during post-deployment checks.
Which one action is essential to confirm that power and cooling are sufficient for the stable operation of NVIDIA DGX H100 systems?
Show answer and explanation
Correct answer: C. Verify that each DGX system is connected to redundant, properly rated PDUs and that all power supplies are reporting nominal input.
A DGX H100 draws a large amount of power and relies on redundant power supplies, so post-deployment validation must confirm that the facility can deliver that power reliably. Checking that each system is fed from redundant, correctly rated PDUs and that every power supply reports nominal input confirms that the system can sustain full load and survive the loss of a feed. Stable power delivery is also the basis for the cooling to work as designed, because the fans and GPUs depend on it, and power supply status can be read from the BMC or with NVSM.
Why the other options are wrong
- A. Disabling devices to reduce heat changes the system configuration and does not confirm that facility power and cooling are adequate.
- B. Overclocking raises power draw and heat and is not a supported validation step for DGX H100 systems.
- D. DGX fans are regulated automatically by the BMC, and forcing constant full speed does not validate cooling capacity.
Question 6
An engineer needs to validate 400G DAC cable signal integrity in a DGX cluster.
Which CVT metric best identifies marginal cables needing replacement?
Show answer and explanation
Correct answer: A. Effective BER > 1.5E-254 during a ≤6-hour monitoring window.
The UFM Cable Validation Tool judges link quality mainly by bit error rate. Effective BER is the error rate that remains after forward error correction, so a value above the acceptance threshold during the monitoring window shows that the cable cannot carry 400G traffic with enough margin. Links that exceed the threshold are flagged as marginal and the cable is reseated or replaced before the cluster goes into production. Measuring BER over a sustained window catches intermittent errors that a short link check would miss, which makes it the most direct indicator of signal integrity for a DAC.
Why the other options are wrong
- B. Matching the transceiver model to a specification confirms compatibility but says nothing about the signal quality of a particular cable.
- C. Temperature variation reflects environmental conditions and is not the metric used to judge the signal integrity of a passive copper cable.
- D. Lane power variance applies to optical transceivers, while a DAC is a passive copper cable whose quality is judged by error rate.
Question 7
A user encounters “permission denied” errors when running GPU-accelerated containers on a Secure Boot-enabled system.
What resolves this?
Show answer and explanation
Correct answer: C. Enroll the MOK and sign NVIDIA kernel modules.
When UEFI Secure Boot is enabled, the Linux kernel loads only modules signed by a trusted key. If the NVIDIA kernel modules are unsigned or signed with a key the firmware does not trust, they fail to load, the GPU device nodes are not usable and containers that request GPUs fail with permission errors. Signing the NVIDIA modules with a Machine Owner Key and enrolling that key through the MOK manager at the next boot lets the kernel load the driver while Secure Boot stays enabled, which restores GPU access for the NVIDIA Container Toolkit.
Why the other options are wrong
- A. SELinux is unrelated to Secure Boot module signature checks, so disabling it does not allow unsigned NVIDIA modules to load.
- B. Removing the NVIDIA runtime takes away GPU support from containers instead of fixing the driver loading failure.
- D. Running Docker with sudo changes user privileges but does not allow the kernel to load unsigned modules under Secure Boot.
Question 8
An InfiniBand administrator needs to run performance benchmarks on new devices added to the fabric.
What tool should be used to check the latency?
Show answer and explanation
Correct answer: B. ib_write_lat
The ib_write_lat utility is part of the perftest package and measures the latency of RDMA write operations between two InfiniBand endpoints. One node runs it as a server and the other connects as a client, and the tool reports minimum, typical and maximum latency, and it can sweep a range of message sizes with the -a option. This makes it the standard way to benchmark the latency of newly added adapters or links and compare the results against expected values. Its companion ib_write_bw measures bandwidth in the same way.
Why the other options are wrong
- A. tcpdump captures IP network packets for analysis and does not benchmark RDMA latency on InfiniBand.
- C. perfmon is a general operating system performance monitor and is not an InfiniBand latency benchmark.
- D. ibdiagnet scans the fabric for errors, topology and link health but does not run point-to-point latency benchmarks.
Question 9
A cluster administrator is preparing to update the firmware on a DGX H100 system, including the GPU tray (baseboard).
What is the correct sequence of steps to perform a safe and successful firmware upgrade?
Show answer and explanation
Correct answer: C. Stop all GPU activity, update and reboot the BMC, update motherboard and tray components, perform a cold reset, and verify completion.
The DGX H100 firmware update procedure starts by stopping all GPU activity, including workloads and services such as DCGM that hold the GPUs, so the components can be flashed safely. The BMC is updated and rebooted first because it manages the remaining updates. The motherboard tray and GPU tray components are then updated, and a cold reset (a full power cycle of the system) activates the new firmware. Finally, the firmware versions are checked to verify that every component reports the expected version.
Why the other options are wrong
- A. Skipping the motherboard and GPU tray updates leaves components on old firmware and mismatched with the rest of the release.
- B. The BMC must be updated and rebooted before the tray components, and the procedure requires a cold reset to activate the new firmware.
- D. A cold reset is used at the end to activate the updated firmware, not at the start before anything has been flashed.
Question 10
One of the nodes in a cluster is not running as fast as the others and the system administrator needs to check the status of the GPUs on that system.
What command should be used?
Show answer and explanation
Correct answer: C. nvidia-smi
The nvidia-smi utility, installed with the NVIDIA driver, reports the state of every GPU in the system: utilization, memory use, temperature, power draw, clock speeds, performance state, ECC errors and running processes. For a node that runs slower than its peers, nvidia-smi quickly shows whether a GPU is missing, throttling because of power or thermal limits, running at reduced clocks or busy with an unexpected process. Its query options also expose the throttle reasons and PCIe link state, which helps narrow down the cause of the slowdown.
Why the other options are wrong
- A. iblinkinfo reports InfiniBand link states across the fabric and does not show GPU status.
- B. Listing PCI devices only confirms that NVIDIA devices are present and gives no utilization, thermal or clock information.
- D. nvidia-gpu-status is not an NVIDIA command.
What people say about the full NCP-AII pack 5.0 from 4
Question 11
For a 48-hour NCCL burn-in test, which parameters ensure sustained fabric stress while detecting silent data corruption?
Show answer and explanation
Correct answer: C. all_reduce_perf -b 8G -e 32G -c 1000 -z 1 -G 1000
In nccl-tests, all_reduce_perf exercises every GPU and link in both directions, and large message sizes from 8G to 32G keep the fabric saturated. The -c flag sets how many iterations have their results checked for correctness, so -c 1000 validates the reduced data on a large number of iterations and exposes silent data corruption. The -z 1 setting makes collectives blocking with a barrier after each iteration, and -G 1000 captures iterations as a CUDA graph and replays it 1000 times, which sustains continuous load. This combination provides both long-duration stress and explicit data validation.
Why the other options are wrong
- A. This command sustains load but leaves correctness checking at its minimal default, so it is weaker at detecting silent data corruption over a long run.
- B. Broadcast moves data in one direction from a root and only sets warmup iterations, so it neither stresses the fabric fully nor adds extended correctness checking.
- D. Specifying only a step factor and GPU count runs a short default sweep with no extended correctness checking or sustained replay for a 48-hour test.
Question 12
An engineer is reimaging a DGX system in a large cluster.
Which method ensures the most efficient and secure remote installation without physical access?
Show answer and explanation
Correct answer: D. Build a software image on Base Command Manager and then reimage.
In a cluster managed by Base Command Manager, DGX nodes are provisioned over the network from software images held on the head node. The administrator builds or updates a software image, assigns it to the node or its category, and triggers a reimage. The node network boots and receives the full image from the head node, with no USB media or physical presence required. Because every node is installed from the same controlled image, the process is repeatable, consistent across the cluster and efficient at scale.
Why the other options are wrong
- A. Booting from a USB drive requires physical access to each system and does not scale across a large cluster.
- B. Skipping image verification removes integrity checks, and raw flashing over SSH is neither a secure nor a supported installation method.
- C. An apt-get upgrade updates packages on the existing installation and does not reimage the system.
Question 13
You are tasked with setting up High Availability (HA) for NVIDIA Base Command Manager (BCM) in a new GPU cluster. The cluster consists of a primary and secondary head node, as well as several compute nodes. The requirements are: automatic failover of BCM services, minimal disruption to workloads, and proper cluster health monitoring during and after installation.
During your BCM HA installation and configuration process, which two of the following actions are mandatory for ensuring a robust and verified HA cluster configuration? (Choose two.)
Show answer and explanation
Correct answer: A, E. After configuration is complete, simulate a failover by stopping BCM services on the active head node to verify that all services are running on the secondary node with no interruption; Assign a floating Virtual IP address that can automatically migrate between the primary and secondary head nodes during failover.
A Base Command Manager high availability pair relies on a shared virtual IP address that always belongs to the active head node. Compute nodes and administrators address cluster services through this floating address, so when the passive head node takes over, the address moves with the services and clients continue without reconfiguration. A configuration is only verified once failover has been exercised: stopping services on the active head node, or forcing a takeover, and confirming that the secondary node assumes all services proves that automatic failover works before the cluster enters production.
Why the other options are wrong
- B. The Base Command Manager HA setup handles cloning and ongoing replication of head node data, so manual directory synchronization is not a required step.
- C. Head node synchronization does not depend on compute nodes being powered on or running jobs.
- D. Separate static addresses without a shared virtual IP force clients to be repointed after failover, which defeats automatic failover.
Question 14
What command sequence is used to identify the exact name of the server that runs as the master SM in a multi-node fabric?
Show answer and explanation
Correct answer: B. sminfo, then smpquery ND
sminfo queries the fabric for the master subnet manager and returns its LID, GUID, priority and state. That LID can then be passed to smpquery with the ND (NodeDescription) attribute, which returns the node description string of the device hosting the subnet manager. On servers the node description normally contains the host name, so this sequence reveals the exact name of the machine running the master SM.
Why the other options are wrong
- A. ibstat shows local adapter and port status, and sminfo alone returns the LID, GUID, priority and state of the master SM rather than the server name.
- C. ibis and ibsim are a fabric inspection shell and a fabric simulator, not the standard way to name the master SM.
- D. smpquery NI returns NodeInfo fields such as GUIDs and port counts, not the node description that holds the host name.
Question 15
After running a 24-hour stress test on a DGX node, the administrator should verify which two key metrics to ensure system stability?
Show answer and explanation
Correct answer: D. No thermal throttling events and consistent GPU utilization ≥95% throughout the test.
A long stress test is meant to prove that the GPUs can sustain full load without degrading. The absence of thermal throttling events shows that cooling and power delivery kept the GPUs within limits for the whole run, so clocks were never reduced to protect the hardware. Consistently high GPU utilization throughout the test shows that the workload kept every GPU busy with no hangs, drops or failed devices. These two checks together confirm both thermal health and sustained performance, which is what stability means for a DGX system under load.
Why the other options are wrong
- A. Total energy use and NVLink bandwidth describe consumption and interconnect performance, and they do not show whether the system stayed stable under sustained load.
- B. SSD write endurance and installed RAM capacity are fixed hardware properties, and a stress run does not validate stability through them.
- C. High CPU usage and container uptime say nothing about GPU health, and GPU-centric workloads do not need to keep the CPUs at high utilization.
Question 16
For an NVIDIA Enterprise AI Factory with 256 GPUs, which storage solution characteristic is most critical to validate during scaling tests?
Show answer and explanation
Correct answer: B. Consistent per-node throughput ≥8 GiB/s.
In a 256-GPU AI factory, many nodes read training data and write checkpoints at the same time, and a single slow node stalls the synchronized training job. The storage characteristic that matters most during scaling tests is therefore sustained, consistent throughput delivered to every node concurrently as the cluster grows. Validating that each node maintains a target bandwidth such as 8 GiB/s under full load confirms that the storage system scales with the GPU count and will keep the GPUs fed, rather than performing well only for isolated clients.
Why the other options are wrong
- A. RAID rebuild time is a resilience consideration, and it does not show whether storage throughput scales with the number of GPUs.
- C. Single-node performance on an idle cluster hides the contention that appears when every node accesses storage at once.
- D. Peak small-block random IOPS does not represent the large, parallel reads and checkpoint writes that dominate AI training workloads.
Question 17
A user wants to restrict a Docker container to use only GPUs 0 and 2.
Which command achieves this?
Show answer and explanation
Correct answer: B. docker run --gpus “device=0,2” nvidia/cuda:12.1-base nvidia-smi
Docker's --gpus option, provided through the NVIDIA Container Toolkit, accepts a device list that limits which GPUs a container can see. Passing --gpus with the value device=0,2 exposes only GPUs 0 and 2, and running nvidia-smi inside the container confirms that just those two devices are visible. NVIDIA documentation writes a list of devices in the form '"device=1,2"', with double quotes inside single quotes, so that Docker reads the comma-separated list as one device value.
Why the other options are wrong
- A. Passing device nodes with --device does not inject the driver libraries and is not the supported way to assign GPUs.
- C. --gpus all exposes every GPU to the container, and arguments to nvidia-smi only change its output.
- D. Without the --gpus option or the NVIDIA runtime set as the default, the environment variable alone exposes no GPUs.
Question 18
Why is it important to provide a large and high-performance local cache (using SSDs configured as RAID-0) for deep learning workloads on DGX systems?
Show answer and explanation
Correct answer: C. A local SSD cache in RAID-0 ensures that most training data is read only once from the network, significantly reducing NFS traffic.
Deep learning training reads the same dataset repeatedly across many epochs. DGX systems provide large NVMe drives configured as a RAID-0 volume so that data fetched from NFS can be cached locally, typically with cachefilesd. After the first epoch most reads come from the fast local cache, so the dataset crosses the network roughly once. This sharply reduces NFS traffic and load on the shared storage server while giving the GPUs a high-bandwidth local data source. RAID-0 is chosen purely for capacity and throughput, since the cache can always be refilled from the network copy.
Why the other options are wrong
- A. A RAID-0 cache volume is ordinary local storage read through the host and does not by itself give GPUs direct file access.
- B. The local cache has nothing to do with tuning NFS server threads or storage reliability on the server side.
- D. RAID-0 stripes data without parity or mirroring, so it provides no redundancy if a drive fails.
Question 19
Which of the following steps are essential components of a recommended DGX cluster installation procedure? (Choose two.)
Show answer and explanation
Correct answer: A, C. Configure networking by validating all interfaces on each node, ensuring proper InfiniBand/Ethernet connectivity, prior to installing cluster software; Group nodes by function during initial setup and assign them to relevant categories in the cluster management tool.
A reliable DGX cluster installation validates the network first: every interface on every node is checked and InfiniBand and Ethernet connectivity is confirmed before cluster software is installed, so that software problems are not confused with cabling or link faults. Nodes are also organized by function during initial setup. In Base Command Manager, nodes are placed into categories (for example, DGX compute nodes versus other node types), and each category carries its software image and settings. This allows consistent provisioning and management of large groups of identical systems.
Why the other options are wrong
- B. Running containerized distributed jobs is a post-installation validation activity, not a step in installing the cluster.
- D. Installing Slurm and configuring node OS images are tasks handled through the cluster manager rather than the essential early steps of validating the network and grouping nodes into categories.
Question 20
After upgrading to HPL-AI 2.0 on a DGX A100 cluster, a 2x performance gain is observed.
Which optimization is primarily responsible for this improvement?
Show answer and explanation
Correct answer: D. MPI-aware GPU communication that reduces CPU bottlenecks and GPU idle time.
HPL-AI, now known as HPL-MxP, solves a dense linear system with low-precision factorization followed by iterative refinement, and at scale its runtime depends heavily on how well inter-GPU communication overlaps with computation. Software improvements that make the MPI layer GPU-aware let data move directly between GPU buffers without staging through host memory, which removes CPU bottlenecks and keeps the GPUs busy instead of waiting on communication. A large performance gain from a software upgrade on the same DGX A100 hardware therefore comes from this communication efficiency rather than from any change to the hardware itself.
Why the other options are wrong
- A. Shrinking the problem size reduces the work the GPUs can overlap and normally lowers the achieved FLOPS, so it cannot explain a doubling of performance.
- B. GPU clock speeds are fixed by the hardware design and power limits, and no firmware update or configuration doubles them.
- C. NVLink bandwidth is a physical property of the GPUs and NVSwitch hardware, and driver updates cannot double it.
Question 21
An engineer needs to validate NVLink Switch functionality on a DGX H100 system with 8 GPUs.
Which NCCL command verifies intra-node NVLink bandwidth?
Show answer and explanation
Correct answer: B. all_reduce_perf -b 8 -e 16G -f2 -g 8 with NCCL_TESTS_SPLIT=“OR 0x7”
NCCL_TESTS_SPLIT divides ranks into separate communicators using an AND, OR, MOD or DIV operation on the rank number. With the setting OR 0x7, every rank on an 8-GPU system maps to the same color, so all eight GPUs of the node form one communicator and the collective is purely intra-node. Running all_reduce_perf with -g 8 across a sweep from 8 bytes to 16G therefore drives traffic among all eight GPUs through the NVLink Switch fabric and reports the intra-node bus bandwidth, which is exactly what validating NVLink Switch functionality on a DGX H100 requires.
Why the other options are wrong
- A. The MOD 2 split creates two communicators that each use every other rank, and with only four GPUs it never exercises all eight GPUs together.
- C. Broadcast is a one-to-all pattern from a single root, so it does not produce the all-to-all NVLink load used to measure intra-node bus bandwidth.
- D. A single-GPU all_reduce performs no inter-GPU communication, so repeating it eight times never measures NVLink bandwidth.
Question 22
As the infrastructure lead for an NVIDIA AI Factory deployment, you have just uploaded the latest supported firmware packages to your DGX system. It is now critical to ensure all hardware components run the new firmware and the DGX returns to full operational capability.
Which sequence best guarantees that all relevant components are correctly running updated firmware according to NVIDIA’s documentation and recommended operational steps?
Show answer and explanation
Correct answer: B. Initiate a cold power cycle on the system to activate firmware for components, reset the BMC using the recommended command, and perform an AC power cycle to ensure EROT and CPLD firmware is activated.
After the firmware packages are applied, different DGX components need different events before they run the new firmware. A cold power cycle of the system activates the firmware on most host and tray components. Resetting the BMC with the documented command makes the management controller load its updated firmware. Some components, including the EROT and CPLD devices, activate only after the system loses AC power, so a full AC power cycle completes the process. Following this order ensures that every component is running the updated firmware before the system returns to service.
Why the other options are wrong
- A. A single AC power cycle followed by a software reset skips the documented BMC reset step of the activation sequence.
- C. An operating system restart does not power cycle the hardware, so most updated firmware remains inactive.
- D. CPLD firmware is activated by removing AC power, not through the management interface after a reboot.
Question 23
You are evaluating the integration of NVIDIA BlueField DPUs into your data center’s storage architecture to optimize AI workloads. The storage solution chosen has incorporated BlueField DPUs to enhance performance and efficiency.
Which of the following benefits directly results from this integration?
Show answer and explanation
Correct answer: C. Reduced CPU load by offloading data processing tasks to DPUs.
BlueField DPUs combine a ConnectX network adapter with Arm cores and hardware accelerators, which lets them run storage, networking and security functions on the DPU instead of on the host. In a storage architecture this includes work such as NVMe over Fabrics processing, storage virtualization, encryption and data movement. Moving those tasks onto the DPU frees host CPU cycles for the AI applications themselves, so reduced CPU load is the benefit that follows directly from the integration. The other choices overstate what the hardware can deliver.
Why the other options are wrong
- A. Offloading can lower latency, but no hardware eliminates latency entirely, because media access and network transit always take time.
- B. NVMe storage speed is determined by the drives and the fabric, so DPU integration does not by itself confer NVMe-class access speeds.
- D. Adding DPUs requires planning for servers, network ports and software, so they do not deliver unlimited scalability without architectural changes.
Question 24
During multi-node HPL burn-in, GPUs show uneven utilization.
Which configuration ensures balanced workload distribution?
Show answer and explanation
Correct answer: B. Set --gpu-affinity and --cpu-affinity to align GPU and NUMA nodes
The HPL benchmark in the NGC HPC-Benchmarks container launches one rank per GPU through its hpl.sh wrapper, which accepts --gpu-affinity and --cpu-affinity options. These map each rank to a specific GPU and to the CPU cores in the NUMA node closest to that GPU. Without this mapping, ranks can share a GPU or run on distant cores, which leaves some GPUs underused and slows the whole run. Setting both affinities so that each GPU is paired with its local NUMA node produces balanced, consistent utilization across all GPUs and nodes.
Why the other options are wrong
- A. HPL_RUN_GEMM_TESTS relates to running GEMM checks and has no effect on how work is distributed across GPUs.
- C. Changing an out-of-core tile size alters blocking for host memory use and does not fix rank to GPU placement.
- D. Enabling NVSHMEM changes the communication path but does not correct uneven mapping of ranks to GPUs and NUMA nodes.
Question 25
A systems administrator is preparing a new DGX server for deployment.
What is the most secure approach to configuring the BMC port during initial setup?
Show answer and explanation
Correct answer: C. Connect the BMC port to a dedicated and firewalled network and change the default admin credentials.
The BMC gives full out-of-band control of the server, including power control, remote console, virtual media and firmware updates, so it is a high-value target. The secure approach is to connect the BMC port only to a dedicated management network protected by firewall rules, so that only authorized administrators can reach it, and to replace the default administrator credentials immediately. Isolation limits who can reach the interface, and changing credentials removes the most commonly exploited weakness. This keeps the BMC available for setup and troubleshooting while protecting it from unauthorized access.
Why the other options are wrong
- A. Exposing the BMC to the internet with default credentials invites immediate compromise of the entire server.
- B. Leaving the BMC disconnected removes out-of-band management during bring-up, which is when remote console and firmware access are most needed.
- D. Placing the BMC on the production network with default credentials exposes management functions to every user and workload on that network.
Question 26
An enterprise IT team has completed the physical installation of an AI factory with a Spectrum-X Ethernet network connected to all GPU servers.
They now need to ensure the environment is ready for scalable AI workload deployment.
What is the recommended sequence of validation steps?
Show answer and explanation
Correct answer: A. Confirm switch and server firmware configuration, test link connectivity and port health, run network benchmarks, validate software stack, then stage AI workload tests.
Validation should build from the bottom of the stack upward. First confirm that switch and server firmware and configuration match the supported versions for Spectrum-X, then test link connectivity and port health so every cable and port is known good. Network benchmarks then verify that the fabric delivers its expected bandwidth and latency. With the network proven, the software stack is validated, and only then are AI workload tests staged. Each layer is confirmed before the next depends on it, which makes any fault easy to isolate.
Why the other options are wrong
- B. Starting with application benchmarks before firmware and links are verified makes it hard to isolate the cause of any problem.
- C. Directory services and access control do not validate the fabric, and putting them first delays checks of network and hardware readiness.
- D. Validating the software stack before links and firmware reverses the order, and OSPF route checks are not part of validating the GPU compute fabric.
Question 27
After ClusterKit reports “GPU-Host latency exceeds threshold”, which NVIDIA diagnostic tool should be used to isolate hardware faults?
Show answer and explanation
Correct answer: C. DCGM Diags dcgmi diag -r 2
When ClusterKit flags excessive GPU to host latency, the next step is to determine whether a GPU or its PCIe path is faulty. DCGM diagnostics are NVIDIA's tool for this, and dcgmi diag -r 2 runs the medium level suite, which adds tests such as PCIe bandwidth and latency checks and GPU memory tests to the quick deployment checks. The results report pass or fail per GPU and per test, which isolates the failing component so it can be serviced.
Why the other options are wrong
- A. nvidia-smi topo -m shows how GPUs and NICs are connected but runs no tests to detect hardware faults.
- B. Repeating ClusterKit for longer only reproduces the symptom and does not identify which component is faulty.
- D. ib_write_bw measures network bandwidth between nodes and does not test the GPU to host path within a node.
Question 28
After initial setup and health checks, the DGX H100 system administrator wants to verify that containers can access GPUs before running production workloads.
Which method is recommended for this validation?
Show answer and explanation
Correct answer: A. sudo docker run --gpus all --rm nvcr.io/nvidia/cuda:12.1.1-base-ubuntu22.04 nvidia-smi
The NVIDIA Container Toolkit exposes GPUs to Docker containers when the --gpus option is used. Running a CUDA base container from NGC with --gpus all and executing nvidia-smi inside it confirms end to end that the runtime is configured, the driver libraries are injected into the container and every GPU is visible. The --rm flag removes the container afterward, so the check leaves nothing behind. DGX documentation uses this same pattern to validate GPU access from containers.
Why the other options are wrong
- B. Running systemctl inside the container says nothing about GPU visibility and normally fails without an init system.
- C. Listing files only shows the container filesystem and gives no information about GPU access.
- D. Without the --gpus option no GPUs are passed into the container, so nvidia-smi cannot show them.
Question 29
A financial services firm is deploying an AI model for fraud detection that requires rapid inference and data retrieval across multiple sites.
Which feature should their storage system prioritize?
Show answer and explanation
Correct answer: D. Multi-protocol data access with low latency.
Fraud detection must score transactions in near real time, so the storage behind the model and its feature data needs low latency to keep inference fast. Because the firm operates across multiple sites with different applications and clients, multi-protocol access such as NFS, SMB and S3 to the same data lets every location and service reach it without copies or conversions. Together these features support rapid inference and consistent data retrieval throughout the organization.
Why the other options are wrong
- A. Tape backup is designed for archival and recovery and is far too slow for real-time inference.
- B. Low-cost HDD storage trades performance for price and cannot meet rapid retrieval requirements.
- C. Moderate speed does not satisfy the rapid inference requirement, since capacity alone does not reduce latency.
Question 30
A customer has just completed the first boot of their DGX system and is prompted to create an administrative user.
What is the correct approach for setting up this user to ensure secure BMC and GRUB access?
Show answer and explanation
Correct answer: A. Create a unique, strong, lower-case username and password that will be used for both BMC and GRUB access, avoiding default or weak credentials.
During DGX OS first boot setup, the administrative user that is created is also used for BMC and GRUB access. NVIDIA requires a lowercase username, and because this one account protects the operating system, the BMC and the boot loader, it should be unique and protected by a strong password. Avoiding default or easily guessed credentials prevents unauthorized changes to boot parameters and out-of-band access to the system through the BMC, which would otherwise expose the whole node.
Why the other options are wrong
- B. Keeping default credentials leaves the BMC and boot loader open to anyone who knows the factory values.
- C. The first boot setup creates a single administrative account that is applied to both BMC and GRUB, not separate accounts.
- D. A predictable username with a simple password is easy to guess and defeats the purpose of securing BMC and GRUB access.
Question 31
A system administrator needs to configure a BlueField DPU and enable RShim on the baseboard management controller (BMC).
Which command should be executed?
Show answer and explanation
Correct answer: B. ipmitool raw 0x32 0x6a 1
The BlueField BMC exposes NVIDIA OEM IPMI commands for managing the RShim interface, which is the channel used to push boot images, access the console and manage the DPU. In that command set, network function 0x32 with command 0x6a controls the RShim service on the BMC, and a data byte of 1 starts it while 0 stops it. Running ipmitool raw 0x32 0x6a 1 against the BMC therefore enables RShim on the BMC, after which the BMC can be used to provision and manage the BlueField DPU.
Why the other options are wrong
- A. No bmc-rshim service unit is used to enable RShim on the BMC, so this systemctl command does not perform the required change.
- C. Restarting the rshim service on a host only reloads the host-side driver, and it does not enable RShim on the BMC.
- D. Copying a file to the rshim boot device installs a BlueField image after RShim is already available, so it does not enable RShim on the BMC.
Question 32
To validate bisectional bandwidth across two racks in a Spectrum-X Ethernet fabric, which NCCL test configuration isolates East-West traffic?
Show answer and explanation
Correct answer: B. NCCL_TESTS_SPLIT=“MOD 2” ./all_reduce_perf -g 8
The nccl-tests NCCL_TESTS_SPLIT variable partitions the ranks into groups that run the same collective in parallel. With MOD 2, the ranks are divided into two groups, each using every other rank, so, when the ranks of the two racks are ordered consecutively, each group includes nodes from both racks. Running all_reduce_perf with eight GPUs per thread (-g 8) in this split forces both operations to send their traffic between the racks at the same time, which loads the East-West paths of the Spectrum-X fabric and measures the bisection bandwidth between the two racks.
Why the other options are wrong
- A. Running without a split and averaging results per rack does not deliberately load the links between the two racks.
- C. DIV 8 groups the ranks node by node, so each operation stays inside a single node and generates no inter-rack traffic.
- D. With eight GPUs per process, OR 0x7 groups blocks of eight consecutive processes together, so it does not create the two interleaved groups needed to load the links between the racks evenly.
Question 33
After configuring HA, the administrator runs cmsh status and notices the secondary head node reports mysql [FAIL].
What is the most likely cause?
Show answer and explanation
Correct answer: D. Network connectivity issues between the primary and secondary head nodes.
In a Base Command Manager high availability setup, the cluster database held in MySQL is replicated between the active and passive head nodes, and the replication runs over the network link between them. When cmsh status shows mysql in a FAIL state on the secondary head node, the usual cause is that the two head nodes cannot communicate reliably, so replication cannot be established or maintained. Restoring connectivity between the head nodes, including the dedicated failover or internal network path, allows the database to synchronize and the check to return to a healthy state.
Why the other options are wrong
- A. GPU drivers play no part in MySQL replication, and head nodes do not need GPUs to run the cluster database.
- B. An expired license affects cluster management features generally and does not produce a MySQL failure on the secondary head node only.
- C. Compute nodes being powered on does not interfere with database replication between the two head nodes.
Question 34
After configuring NGC CLI with ngc config set, a user receives “Authenticated failed” errors when pulling containers.
What step was most likely omitted?
Show answer and explanation
Correct answer: B. Entering the API key during ngc config set or storing it in ~/.ngc/config.
The ngc config set command prompts for the NGC API key along with the output format, organization and team, and saves the values in the configuration file under ~/.ngc. Every authenticated request to NGC uses this key, so if it is skipped or not stored, the user cannot authenticate and container pulls fail. For pulls through Docker, the same API key is also used as the password for docker login to nvcr.io with the username $oauthtoken.
Why the other options are wrong
- A. The method used to install the CLI has no effect on authentication, which depends on the API key.
- C. Restarting Docker does not supply credentials, so authentication still fails without an API key.
- D. The format type controls only how the CLI displays output and is not involved in authentication.
Question 35
A customer is designing an AI Factory for enterprise-scale deployments and wants to ensure redundancy and load balancing for the management and storage networks.
Which feature should be implemented on the Ethernet switches?
Show answer and explanation
Correct answer: B. MLAG for bonded interfaces across redundant switches.
MLAG (multi-chassis link aggregation) lets two Ethernet switches act as a single logical peer, so a server can bond one link to each switch into one LAG. Both links carry traffic at the same time, which balances load, and if a switch or link fails the bond continues on the remaining path. NVIDIA reference architectures use MLAG on the management and storage switches for this reason, removing the single point of failure while using all available bandwidth.
Why the other options are wrong
- A. A single switch is a single point of failure and provides neither redundancy nor load balancing.
- C. Spanning tree blocks redundant paths to prevent loops, so it gives failover but no active load balancing.
- D. Unmanaged switches without VLANs remove segmentation and cannot provide aggregated redundant links.
Question 36
You are tasked with updating both NVIDIA GPU drivers and DOCA drivers on a set of servers used for AI workloads. The environment previously had an older driver stack and custom kernel modules.
What is the most important step to successfully upgrade the drivers without causing conflicts?
Show answer and explanation
Correct answer: B. Uninstall all existing GPU and DOCA-related drivers and associated kernel modules before the new install
When a system carries an older driver stack and custom kernel modules, leftover packages and modules can conflict with the new release. Old modules may load in place of the new ones, and mismatched libraries can break the link between the GPU driver and the DOCA networking stack. Removing the existing GPU driver, DOCA or MLNX_OFED packages and any associated kernel modules before installing gives the new versions a clean base, so the correct modules are built and loaded for the running kernel and the two stacks interoperate as intended.
Why the other options are wrong
- A. Leaving the old DOCA and OFED stack in place risks incompatibility with the new GPU driver even when hardware is detected.
- C. Only one version of a kernel driver can be loaded at a time, so keeping both installed invites conflicts.
- D. A new installation does not reliably remove old packages and custom modules, so checking the version afterward does not prevent conflicts.
Question 37
ClusterKit’s NCCL bandwidth test shows 350 GB/s on a 400G InfiniBand fabric.
How should this result be interpreted?
Show answer and explanation
Correct answer: B. Optimal performance, indicating healthy fabric and GPUDirect RDMA.
ClusterKit reports NCCL bus bandwidth for each node, which aggregates traffic across all of the node's GPUs and their network adapters. On a DGX-class node with eight 400 Gb/s InfiniBand ports, the raw aggregate is about 400 GB/s, and protocol overhead means well-tuned systems reach roughly 350 GB/s or slightly more. A result of 350 GB/s is therefore in the expected range and shows that the fabric links are healthy and that GPUDirect RDMA is moving data directly between GPU memory and the adapters without staging through host memory.
Why the other options are wrong
- A. CPU stress does not exercise the GPU-to-network data path, so rerunning with it adds nothing to the validation of fabric bandwidth.
- C. The result is already in the expected range, and FEC settings govern error correction on links rather than serving as a bandwidth tuning knob.
- D. HDR InfiniBand runs at 200 Gb/s per port, so it does not apply to a 400G fabric, and the stated threshold is not a valid expectation.
Question 38
During cluster validation, the Cable Validation Tool (CVT) reports “Underperforming (BER)” for an InfiniBand link.
Which BER thresholds indicate a critical signal quality issue requiring cable replacement?
Show answer and explanation
Correct answer: A. Raw BER > 1e-6 or Effective BER > 1.5E-254 for ≤6hr measurements
The Cable Validation Tool judges link signal quality from bit error rate measurements taken over a defined measurement window. Raw BER is the error rate before forward error correction, and effective BER is the rate remaining after FEC. A raw BER above 1e-6, or an effective BER above the tool's limit for measurements of up to six hours, means the link cannot reliably keep errors under control, so CVT reports it as underperforming and the cable or transceiver should be replaced.
Why the other options are wrong
- B. This describes an error count condition during early link uptime rather than the raw and effective BER rate thresholds the question asks for.
- C. Transceiver temperature is a thermal alarm and is not a BER threshold for signal quality.
- D. Receive power differences between lanes are an optical power check and not the BER criterion used for this result.
Question 39
After Spectrum-X fabric deployment, NCCL tests show intermittent latency spikes.
Which network condition most severely impacts East-West bandwidth?
Show answer and explanation
Correct answer: C. Packet loss >0.001% causing NCCL pipeline stalls.
RDMA traffic over RoCE, which carries NCCL collectives on Spectrum-X, is designed for a lossless fabric. When packets are dropped, the RDMA transport must detect the loss and retransmit, and because collective operations proceed in lockstep across all participating GPUs, a single delayed transfer stalls the entire pipeline. Even a very small loss rate therefore produces latency spikes and a large drop in effective east-west bandwidth. Eliminating drops through correct congestion control and lossless configuration is the priority when NCCL tests show this behavior.
Why the other options are wrong
- A. Transceiver firmware mismatches tend to cause link instability or failure rather than being the main driver of intermittent bandwidth loss.
- B. Seventy percent port utilization during a bandwidth test is expected load and does not by itself degrade east-west performance.
- D. Jitter below 5 microseconds with consistent latency describes a healthy fabric, not a cause of degradation.
Question 40
You are expanding a DGX-based deep learning cluster to train on large, high-resolution images that cannot fit into local cache. Multiple nodes will access this data concurrently and require high performance.
Which storage and networking solution best meets these requirements?
Show answer and explanation
Correct answer: D. Deploy a high-performance parallel filesystem across InfiniBand or 40/100GbE, ensuring at least 3GB/s per node and scalable aggregate bandwidth for all cluster workloads.
When a training dataset is too large for local cache and many DGX nodes read it at once, the storage tier must provide high per-node throughput and aggregate bandwidth that grows with the cluster. A high-performance parallel filesystem spreads data and metadata across many servers and serves clients over InfiniBand or high-speed Ethernet, so each node can sustain several gigabytes per second while total bandwidth scales with added storage servers. This keeps the GPUs fed with data and prevents I/O from becoming the training bottleneck.
Why the other options are wrong
- A. A single NFS server on 10GbE cannot deliver multi-gigabyte-per-second reads to many nodes concurrently and quickly becomes the bottleneck.
- B. The dataset is too large to fit in local cache, so enlarging local RAID-0 does not remove the dependence on shared storage performance.
- C. General-purpose object storage is not built for high-throughput, low-latency concurrent reads by many training nodes.
Question 41
A system administrator noticed a failure on a DGX H100 server. After a reboot, only the BMC is available.
What could be the reason for this behavior?
Show answer and explanation
Correct answer: A. There are more than two failed power supplies.
The DGX H100 has six power supply units in a 4+2 redundant configuration, so it runs normally with up to two failed units, and with three failed it keeps running at reduced performance. The BMC runs from standby power and stays reachable even when the host cannot start. NVIDIA documents that the system boots only while at least three supplies are operational; with fewer than three operational, only the BMC is available. Of the listed causes, only failed power supplies keep the host from powering on, and the BMC power and sensor views, together with its event logs, identify which supplies have failed.
Why the other options are wrong
- B. A failed boot disk still lets the system power on and reach firmware, where the boot failure is visible on the console, so the host is not limited to the BMC.
- C. A network card without link affects connectivity only, and the host still powers on and boots normally.
- D. Failed GPUs are reported by the operating system and health tools after boot, and they do not stop the host from powering on.
Question 42
A media company is developing an AI platform for video content analysis that requires storing and processing large volumes of unstructured video data. The platform must support high throughput for data ingestion and provide efficient access for real-time analytics.
Given these requirements, which storage strategy should the company implement?
Show answer and explanation
Correct answer: B. Object storage for scalability and metadata management
Large volumes of unstructured video fit naturally in object storage, which uses a flat namespace that scales horizontally to petabytes and billions of objects. Each object carries rich, customizable metadata, which helps analytics pipelines locate and classify content such as scenes, timestamps or tags. Object stores accessed through S3-compatible APIs also support highly parallel ingestion and retrieval, delivering the aggregate throughput a video analytics platform needs as it grows.
Why the other options are wrong
- A. Block storage suits databases and virtual machine disks but lacks the scale and metadata model needed for massive unstructured video collections.
- C. Tape is suited to long-term archives and cannot provide the fast access required for real-time analytics.
- D. Hierarchical file storage becomes harder to scale and manage at very large file counts and offers limited custom metadata.
That was 42 of 170.
The full NVIDIA NCP-AII pack has all 170 questions, each with the answer, the explanation and why the other options are wrong, plus a questions-only copy for timed runs. US$39, paid once, with free monthly updates and a pass-or-your-money-back guarantee.
Not ready yet? Get the first 85 free as one PDF by email. Get the Half Pack →