I85PREP
⌘ K
NVIDIA NCP-AII study materials
NVIDIA NCP-AII study resources

NVIDIA NCP-AII practice questions

One question at a time. Build your NVIDIA NCP-AII understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1System and Server Hardware Installation

When racking a GPU server that ships with redundant hot-swap power supplies, which practice best supports both cooling and serviceability in a dense NVIDIA GPU node?

  1. Connect both supplies to the same PDU outlet using a Y-splitter cable
  2. Leave PSU bays empty and rely on the rack-level bus bar for power
  3. Populate all PSU bays and connect each supply to a separate PDU feed so the node survives a single branch failure Correct answer.
  4. Install a single high-wattage PSU and leave the second bay empty to reduce heat output

Explanation

Redundant hot-swap supplies are designed to be populated and fed from independent power paths so that loss of one PDU branch or supply does not drop the node. Feeding both supplies from one outlet defeats redundancy, and leaving bays empty removes both capacity and fault tolerance.

Question 2System and Server Hardware Installation

A technician is installing an NVIDIA data-center GPU card into a PCIe slot. Which combination of connections must be completed before the card can be fully powered and managed?

  1. PCIe slot plus any auxiliary GPU power connectors, with the card seated and bracket secured Correct answer.
  2. PCIe slot only, since all data-center GPUs draw power solely from the slot
  3. An external power brick connected to the GPU's NVLink port
  4. Auxiliary power connectors only, with the card left loosely seated in the slot

Explanation

High-power data-center GPUs draw more than the PCIe slot can supply, so auxiliary power cables must be connected in addition to proper slot seating. The bracket must also be secured so the card stays seated and the chassis retains airflow and EMI integrity.

Question 3System and Server Hardware Installation

Which memory population rule should be followed when installing DIMMs for a dual-socket server that will host NVIDIA GPUs?

  1. Install DIMMs in any order, since the memory controller auto-balances channels at boot
  2. Populate only one DIMM per socket to minimize latency regardless of channel count
  3. Fill all slots on socket 0 first, then populate socket 1 only if errors appear
  4. Follow the platform population table, balancing DIMMs evenly across memory channels on each socket Correct answer.

Explanation

Server memory controllers gain bandwidth from interleaving across channels, so DIMMs must be placed per the vendor population table and balanced per socket. Unbalanced or arbitrary placement can reduce bandwidth and cause population errors that block boot.

Question 4System and Server Hardware Installation

When cabling NVLink bridges or GPU-to-GPU interconnect hardware inside a server, what is the primary installation concern?

  1. Routing and seating the interconnect per the platform guide so links train at full width Correct answer.
  2. Connecting interconnect cables while the system is powered on
  3. Using the longest available cables to simplify routing
  4. Coiling interconnect cables tightly around the CPU heatsink

Explanation

GPU interconnect hardware must be seated and routed exactly as the platform guide specifies so all lanes train and the topology is recognized. Excess cable length or tight bends can degrade signal integrity and cause link errors.

Question 5System and Server Hardware Installation

What is the purpose of installing the server rails and cable management arm before sliding a heavy GPU node into the rack?

  1. To support the node's weight and manage cable slack so the node can be serviced without strain Correct answer.
  2. To provide electrical grounding for the GPU cards
  3. To lock the node permanently so it cannot be removed
  4. To increase airflow through the front of the chassis

Explanation

Rails carry the weight of a dense GPU chassis and keep it aligned in the rack, while the cable management arm takes up slack as the node slides. This allows the node to be pulled out for service without stressing connectors or cables.

Question 6System and Server Hardware Installation

After physically installing a new NVIDIA GPU into a server, which step confirms the operating system can see the device before driver-level validation?

  1. Reinstall the operating system
  2. Replace the system fans
  3. Query the PCIe bus enumeration to confirm the card's vendor and device IDs appear Correct answer.
  4. Run a full GPU compute benchmark immediately

Explanation

PCIe enumeration tools show whether the card is detected on the bus with the expected vendor and device identifiers. Confirming enumeration first isolates seating and power problems from driver or software issues.

Question 7System and Server Hardware Installation

Why should empty drive bays and unused PCIe slots on a GPU server be covered with blanking panels or filler brackets?

  1. To prevent the GPU from drawing too much power
  2. To maintain designed front-to-back airflow and keep GPU intake temperatures within spec Correct answer.
  3. To allow hot air to recirculate for even cooling
  4. To increase the number of available PCIe lanes

Explanation

Blanking panels preserve the chassis pressure differential so cool intake air is pulled across components instead of bypassing them through open bays. Recirculation through open slots raises GPU intake temperatures and can trigger thermal throttling.

Question 8System and Server Hardware Installation

A server will host multiple GPUs that require direct GPU-to-GPU communication for training. Which hardware consideration most affects whether the intended high-bandwidth topology is achievable?

  1. The number of USB ports on the rear panel
  2. The size of the operating system swap partition
  3. The color of the chassis front bezel
  4. Whether the platform supports the required GPU interconnect and the cards are placed in the designated slots Correct answer.

Explanation

High-bandwidth GPU-to-GPU topologies depend on platform support for the specific interconnect and on installing cards in the slots the design ties together. Slot choice and platform capability determine whether full-bandwidth links can be formed.

Question 9System and Server Hardware Installation

During installation of a multi-GPU server, a technician notices the platform requires a specific slot for the primary GPU used for console output. Why does slot placement matter here?

  1. Only the primary slot supplies auxiliary power to the card
  2. Slot position changes the GPU's clock speed permanently
  3. Firmware may assign the primary display or boot GPU based on slot position, affecting POST visibility Correct answer.
  4. Cards in other slots cannot run compute workloads

Explanation

Firmware often selects a primary GPU by slot or bus order, which determines where POST and console output appear. Placing the intended card in the documented slot ensures the system initializes and displays as expected.

Question 10System and Server Hardware Installation

A rack contains several GPU nodes drawing high current on 240 V feeds. Which installation practice reduces the risk of overloading a branch circuit?

  1. Use extension cords to reach distant outlets
  2. Plug all nodes into one PDU for simpler cable management
  3. Balance node power cords across available PDU circuits and verify load against the circuit rating Correct answer.
  4. Disable PSU redundancy to halve the current draw

Explanation

Spreading cords across circuits keeps each branch within its rated capacity and preserves redundancy when a circuit trips. Verifying calculated load against the breaker rating prevents nuisance trips and overheating.

Question 11System and Server Hardware Installation

A newly installed GPU is not detected after the server boots. Which troubleshooting sequence is most appropriate?

  1. Immediately replace the motherboard
  2. Update the GPU benchmark score database
  3. Power down, reseat the card and auxiliary power, verify slot population rules, then re-check enumeration Correct answer.
  4. Ignore the issue since GPUs are optional

Explanation

A missing GPU is most often caused by poor seating, unconnected auxiliary power, or a slot that is disabled by population rules. Reseating and verifying power and slot configuration before replacing major parts is the efficient path.

Question 12System and Server Hardware Installation

When installing a high-TDP GPU that uses a 12VHPWR-style connector, what should the technician verify about the cable?

  1. The connector is only partially inserted to allow airflow
  2. Any spare connector on the cable is left dangling inside the chassis
  3. The connector is fully inserted until latched and the cable is rated for the GPU's power requirement Correct answer.
  4. The cable is bent sharply at the connector to save space

Explanation

High-power GPU connectors must be fully seated and latched, and the cable must be rated for the card's power draw to avoid overheating and contact damage. Sharp bends at the connector or partial insertion increase resistance and risk.

Question 13System and Server Hardware Installation

A server platform requires that GPUs be installed in pairs across specific CPU sockets to enable balanced PCIe bandwidth. What is the best way to confirm correct installation?

  1. Assume any open slot provides full bandwidth
  2. Review the platform's slot-to-CPU mapping and verify each GPU's link width and affinity after install Correct answer.
  3. Rely on the operating system to reassign PCIe lanes automatically
  4. Install all GPUs in the slots closest to the front of the chassis

Explanation

PCIe slots are wired to specific CPU sockets, and the platform guide documents which slots provide full-width links. After installation, checking negotiated link width and NUMA affinity confirms the GPUs are attached as intended.

Question 14System and Server Hardware Installation

During a rack installation, a technician must ensure the GPU node's hot aisle and cold aisle orientation is correct. Which action supports this?

  1. Seal both the front and rear of the chassis
  2. Confirm the chassis intake faces the cold aisle and exhaust faces the hot aisle per airflow direction Correct answer.
  3. Orient the node so its rear faces the cold aisle
  4. Reverse all fan modules to pull air from the rear

Explanation

Servers are designed for front-to-back airflow, so the intake must face the cold aisle and exhaust must discharge into the hot aisle. Reversing orientation or fans breaks the thermal design and raises component temperatures.

Question 15System and Server Hardware Installation

A GPU node uses liquid cooling for the highest-TDP accelerators. Which installation step is most critical before powering on?

  1. Fill the loop with tap water to save cost
  2. Disconnect the leak sensor to avoid false alarms
  3. Power on immediately to check for leaks
  4. Verify coolant loop connections are secure and leak-tested per the vendor procedure Correct answer.

Explanation

Liquid-cooled accelerators require secure, leak-tested loop connections before energizing to prevent coolant from reaching electronics. Vendor procedures specify the correct coolant and test steps.

Question 16System and Server Hardware Installation

Which statement correctly describes how ECC memory relates to GPU server hardware installation?

  1. ECC is enabled only when the GPU driver is installed
  2. ECC DIMMs must be populated per the platform's supported configuration to protect against memory errors Correct answer.
  3. ECC DIMMs can be mixed with non-ECC DIMMs in any slot without issue
  4. ECC memory eliminates the need for airflow management

Explanation

ECC memory protects against single-bit errors, but only when DIMMs are installed in a supported configuration documented by the platform. Mixing ECC and non-ECC modules or unsupported populations can prevent ECC operation or block boot.

Question 17System and Server Hardware Installation

A technician is installing a GPU riser assembly. Which practice best preserves signal integrity for the PCIe links?

  1. Fold the ribbon cable sharply to fit around a heatsink
  2. Seat the riser fully in its slot and route the ribbon or cable without sharp bends or tension Correct answer.
  3. Route the ribbon cable across fan intakes to save space
  4. Leave the riser partially inserted to reduce strain

Explanation

Riser assemblies carry high-speed PCIe signals that are sensitive to impedance changes from sharp bends and poor seating. Full insertion and gentle routing preserve signal integrity and link training.

Question 18System and Server Hardware Installation

During POST, a GPU server reports a memory training failure after new DIMMs were installed. What is the most likely cause?

  1. The GPU driver is outdated
  2. The operating system swap file is too small
  3. The rack cable management arm is misaligned
  4. DIMMs were populated in an unsupported order or a module is not seated correctly Correct answer.

Explanation

Memory training failures at POST typically stem from incorrect population order, mismatched modules, or a DIMM that is not fully seated. Rechecking against the population table and reseating resolves most cases.

Question 19System and Server Hardware Installation

A server must support GPUs that require more power than a single PSU can deliver. Which installation approach is correct?

  1. Use a lower-wattage PSU to reduce heat
  2. Populate the number of PSUs the platform specifies for the GPU configuration and connect them to adequate circuits Correct answer.
  3. Install one PSU and rely on the GPU's slot power
  4. Connect PSUs to a single shared outlet regardless of rating

Explanation

GPU configurations have documented PSU requirements to supply peak power and maintain redundancy. Populating the specified supplies and feeding them from adequately rated circuits ensures stable operation.

Question 20System and Server Hardware Installation

When installing multiple GPU nodes in a rack, why is it important to record each node's slot position and cabling in an inventory?

  1. It is required to enable GPU clock boost
  2. It supports service identification, firmware matching, and troubleshooting of interconnect and power paths Correct answer.
  3. It removes the need for blanking panels
  4. It allows the rack to exceed its weight limit safely

Explanation

Accurate inventory records let technicians locate a specific node, match firmware and configuration, and trace power and interconnect cabling during service. This reduces downtime and misdiagnosis in dense racks.

Question 21System and Server Hardware Installation

A GPU server will be installed in a rack with limited depth. Which specification must be checked to confirm the node will fit and be serviceable?

  1. The color of the front panel
  2. Only the chassis width
  3. The number of GPU fans
  4. Chassis depth including cable management clearance versus usable rack depth Correct answer.

Explanation

Rack depth must accommodate the chassis plus clearance for rear cabling and the cable management arm so the node can slide for service. Checking only width or front-panel features misses the limiting dimension.

Question 22System and Server Hardware Installation

A technician installs a GPU into a slot that is electrically x8 while the card supports x16. What is the expected outcome?

  1. The card runs at x16 because it negotiates upward
  2. The slot provides more power to compensate
  3. The card refuses to initialize at all
  4. The card trains at x8, reducing available bandwidth to and from the GPU Correct answer.

Explanation

PCIe links train at the highest width supported by both the slot and the card, so an x8 slot yields an x8 link. The GPU still functions but has less host bandwidth, which can bottleneck data transfer.

Question 23System and Server Hardware Installation

A rack-level power budget shows a new GPU node would push a circuit past 80 percent of its breaker rating. What is the most appropriate action before installation?

  1. Remove the breaker to increase capacity
  2. Proceed anyway since breakers tolerate brief overloads
  3. Redistribute the node's supplies across circuits or provision an additional circuit to stay within safe loading Correct answer.
  4. Reduce PSU redundancy to lower the calculated load

Explanation

Sustained loading near or above a breaker's rating causes nuisance trips and heat, so power should be redistributed or additional capacity provisioned. Removing protective devices or ignoring the budget creates a safety hazard.

Question 24System and Server Hardware Installation

After installing several GPUs, a technician observes that one GPU reports a narrower PCIe link than the others in the same model of server. Which investigation is most diagnostic?

  1. Increase the fan speed
  2. Replace all GPUs at once
  3. Reinstall the operating system
  4. Compare the slot-to-CPU mapping and check for a riser or seating fault on the affected slot Correct answer.

Explanation

A single narrower link points to slot wiring, a riser fault, or poor seating rather than a platform-wide limit. Comparing the affected slot against the documented mapping and reseating isolates the cause efficiently.

Question 25System and Server Hardware Installation

A GPU node exhibits intermittent link errors on its GPU interconnect after installation. Which factor is most likely to be the root cause?

  1. The rack door being open
  2. The operating system using too much swap space
  3. The GPU driver version being too new
  4. Improperly seated or excessively bent interconnect cabling degrading signal integrity Correct answer.

Explanation

Intermittent interconnect errors often trace to physical issues such as poor seating, bent cables, or damaged connectors that degrade signal integrity. Reseating and inspecting the interconnect hardware per the platform guide is the first diagnostic step.

Question 26System and Server Hardware Installation

When preparing a server chassis to accept a high-TDP data-center GPU, which environmental requirement is most directly tied to the GPU's thermal design power and airflow path?

  1. Installing the node in the coldest available rack position regardless of airflow direction
  2. Maintaining the chassis inlet air temperature within the vendor's supported operating range Correct answer.
  3. Increasing the fan speed to maximum in firmware before the GPU is seated
  4. Selecting a rack with the greatest number of unused U-space above the node

Explanation

GPU thermal design depends on a defined inlet air temperature and airflow volume; exceeding the supported inlet range causes throttling or thermal shutdown. Vendor installation guides specify this inlet range, so the installer must confirm the datacenter cold aisle meets it. Rack position and fan presets cannot compensate for an out-of-spec inlet temperature.

Question 27System and Server Hardware Installation

A technician is installing a GPU node whose power supply accepts both 120 V and 240 V input. Which installation factor most affects whether the node can deliver full GPU power?

  1. The number of network uplinks provisioned for the node
  2. The color coding of the power distribution unit outlets
  3. The input voltage and the resulting current available per feed Correct answer.
  4. The length of the rack's grounding strap

Explanation

PSU output capability and feed current depend on input voltage; at lower voltage a given circuit delivers less power and may not support full GPU load. The installer must confirm the feed voltage and per-circuit current match the node's power budget. Outlet color, network uplinks, and grounding strap length do not determine available power.

Question 28System and Server Hardware Installation

During a rack build, a technician must install a GPU node that ships with a rail kit supporting both square-hole and threaded-hole racks. Which step best ensures the node slides in and locks correctly at the intended U position?

  1. Attach the rails to the node first, then mark the rack holes by eye
  2. Install the node from the rear of the rack to avoid cable interference
  3. Set the rail depth to match the rack and engage the correct mounting pins for the hole type Correct answer.
  4. Torque all rail screws to the maximum value the driver allows

Explanation

Rail kits have adjustable depth and interchangeable or selectable mounting features for square versus threaded holes; matching both lets the rail sit flush and lock at the correct U. Eyeballing holes risks misalignment and uneven support. Over-torquing and rear insertion are not correct installation practices.

Question 29System and Server Hardware Installation

An installer must connect a GPU node's auxiliary power cables to a power distribution unit. Which practice best reduces the risk of a single feed failure taking down multiple GPUs in the same node?

  1. Daisy-chaining both PSUs of the node from one PDU outlet strip
  2. Spreading the node's redundant PSU inputs across independent PDU feeds Correct answer.
  3. Using the shortest possible cables so both PSUs share one circuit
  4. Connecting all nodes on the rack to the same PDU for simpler cable management

Explanation

Redundant power supplies provide protection only when each input is fed from an independent source such as separate PDUs or circuits. Spreading the PSU inputs across independent feeds means a single feed or PDU failure leaves the node powered. Sharing one strip or circuit defeats the redundancy design.

Question 30System and Server Hardware Installation

A GPU node is being installed in a rack that already contains several high-power nodes. Which combination of actions best verifies the installation meets both electrical and thermal design limits before the node is powered on?

  1. Confirm the node's serial number is recorded and the rack door closes without contact
  2. Confirm the management port links at the expected speed and the fans spin up on power
  3. Confirm the node's firmware is the newest release and the rack has spare U-space
  4. Confirm the circuit loading stays within the derated breaker limit and the cold-aisle inlet temperature is within the node's supported range Correct answer.

Explanation

Commissioning a high-power GPU node requires checking both the electrical budget against derated circuit capacity and the thermal environment against the supported inlet range. These two limits determine whether the node can run at full load without tripping breakers or throttling. The other checks are useful records or post-power observations but do not validate design limits.

Question 31GPU and Accelerator Configuration

When you install a data-center GPU into a PCIe slot, what does the auxiliary power connector on the top edge of the card provide?

  1. A management sideband channel for out-of-band monitoring
  2. Supplemental 12 V power beyond the slot's rated wattage Correct answer.
  3. The reference clock used to synchronize PCIe lanes
  4. A redundant data path for GPU-to-CPU transfers

Explanation

High-power PCIe accelerators draw more than the slot can safely supply, so the 8-pin or 12-pin auxiliary connector feeds additional 12 V rails directly from the PSU. The slot itself is limited to roughly 75 W, which is insufficient for modern data-center GPUs under load.

Question 32GPU and Accelerator Configuration

A server reports that a newly installed A100 is present on the PCIe bus but is not usable by CUDA. Which check most directly confirms whether the NVIDIA kernel driver bound to the device?

  1. Run ip link to verify the management network state
  2. Run lspci -nn to confirm the vendor and device IDs
  3. Run dmidecode to read the system memory configuration
  4. Run nvidia-smi to read the driver version and GPU list Correct answer.

Explanation

nvidia-smi communicates through the NVIDIA kernel driver, so a populated GPU table with a driver version proves the driver loaded and bound. lspci only shows that the PCIe device enumerated, which can succeed even when no driver is attached.

Question 33GPU and Accelerator Configuration

Which mechanism lets multiple GPUs in one node exchange data directly without staging through host memory over PCIe?

  1. A shared memory segment mapped into each process
  2. The baseboard management controller's sideband bus
  3. A loopback network interface bound to each GPU
  4. NVLink peer-to-peer links between the GPUs Correct answer.

Explanation

NVLink provides high-bandwidth direct interconnects between GPUs so that peer-to-peer transfers bypass the host PCIe path. When NVLink is unavailable, P2P over PCIe may still work but at much lower bandwidth and with topology constraints.

Question 34GPU and Accelerator Configuration

You want to restrict a containerized workload to see only GPUs 2 and 3 of a node. Which environment variable does the NVIDIA Container Toolkit honor for this selection?

  1. NVIDIA_VISIBLE_DEVICES set to 2,3 Correct answer.
  2. CUDA_VISIBLE_DEVICES set inside the image build
  3. NVIDIA_DRIVER_CAPABILITIES set to compute
  4. LD_LIBRARY_PATH pointing at the driver directory

Explanation

The NVIDIA Container Toolkit reads NVIDIA_VISIBLE_DEVICES at container start to decide which devices to expose, using indices or GPU UUIDs. Setting it to 2,3 injects only those two devices into the container's runtime configuration.

Question 35GPU and Accelerator Configuration

Which command reports the current GPU clock, memory clock, temperature, and power draw for each installed accelerator?

  1. nvidia-smi --query-gpu=name --format=csv
  2. nvidia-smi --gpu-reset
  3. nvidia-smi -q -d CLOCK,TEMPERATURE,POWER Correct answer.
  4. nvidia-smi -L

Explanation

The -q flag requests detailed per-device information and -d selects which sections to include, such as CLOCK, TEMPERATURE, and POWER. This yields the full telemetry block rather than a short summary.

Question 36GPU and Accelerator Configuration

A GPU shows an Xid error in the system log during a training run. What does an Xid message represent?

  1. A firmware version mismatch reported by the BMC
  2. A CUDA compiler warning emitted during kernel build
  3. A driver-reported GPU error code with a specific failure category Correct answer.
  4. A PCIe link speed downgrade negotiated at boot

Explanation

Xid messages are generated by the NVIDIA driver and carry a numeric code that categorizes the fault, such as memory errors or fallen-off-the-bus conditions. Admins use the code to decide between reset, reseat, or RMA actions.

Question 37GPU and Accelerator Configuration

Which setting enables NVIDIA persistence mode on a GPU so the driver stays initialized between jobs?

  1. nvidia-smi -pl 250
  2. nvidia-smi -lgc 1410
  3. nvidia-smi -pm 1 Correct answer.
  4. nvidia-smi -c EXCLUSIVE_PROCESS

Explanation

The -pm flag toggles persistence mode, and a value of 1 keeps the driver loaded so initialization latency between jobs is avoided. It is commonly enabled on shared compute nodes to reduce first-kernel startup overhead.

Question 38GPU and Accelerator Configuration

What is the purpose of MIG on supported data-center GPUs?

  1. Partitioning one GPU into isolated compute and memory slices Correct answer.
  2. Encrypting GPU memory contents at rest
  3. Extending GPU memory capacity into host DRAM
  4. Multiplexing several GPUs into a single larger CUDA device

Explanation

Multi-Instance GPU carves a physical GPU into independent instances, each with dedicated SMs, memory, and cache, so tenants are isolated. This improves utilization for smaller workloads that do not need the whole device.

Question 39GPU and Accelerator Configuration

A node with eight GPUs must run one training job that needs all eight devices. Which configuration lets the job see every GPU while preventing other processes from sharing them?

  1. Set compute mode to EXCLUSIVE_PROCESS and launch one process per GPU Correct answer.
  2. Enable persistence mode and launch a single process
  3. Set the application clocks to the maximum supported values
  4. Enable MIG mode with one instance per GPU

Explanation

EXCLUSIVE_PROCESS gives each GPU to a single process context, which prevents other processes from attaching, while a distributed launcher starts one rank per device. This is the standard pattern for multi-GPU training with strict device ownership.

Question 40GPU and Accelerator Configuration

You need to confirm the NUMA affinity between a GPU and the CPU socket it is attached to so that you can pin host threads correctly. Which tool output gives this mapping most directly?

  1. nvidia-smi topo -m Correct answer.
  2. nvidia-smi nvlink -s
  3. nvidia-smi --query-gpu=uuid
  4. nvidia-smi -q -d MEMORY

Explanation

The topo -m view prints the GPU interconnect matrix and annotates CPU affinity, showing which socket each GPU sits closest to. That mapping is what you need to bind host threads to the matching NUMA node.

Question 41GPU and Accelerator Configuration

A GPU is shared by several users and needs a hard cap of 200 W to fit the rack power budget. Which command applies that limit at runtime?

  1. nvidia-smi -r
  2. nvidia-smi -ac 5001,1410
  3. nvidia-smi -pl 200 Correct answer.
  4. nvidia-smi -pm 1

Explanation

The -pl flag sets the board power limit in watts, provided the requested value is within the card's allowed range. The change takes effect immediately and can be made persistent with -pm.

Question 42GPU and Accelerator Configuration

A monitoring stack must collect GPU utilization and memory usage every few seconds without invoking a heavyweight query each time. Which mechanism is designed for this?

  1. NVIDIA Data Center GPU Manager in query or metrics mode Correct answer.
  2. Parsing the Xid lines from the kernel log
  3. Direct reads of /proc/driver/nvidia/version
  4. Repeated nvidia-smi -q calls from cron

Explanation

DCGM exposes GPU health and utilization metrics through a resident service and can be scraped by Prometheus or queried via dcgmi. It is built for continuous collection and avoids the overhead of spawning nvidia-smi repeatedly.

Question 43GPU and Accelerator Configuration

A GPU appears in lspci but nvidia-smi reports that it cannot communicate with the driver. What is the most likely cause?

  1. The GPU's ECC memory has corrected errors
  2. The PCIe slot is running at Gen3 instead of Gen4
  3. The NVIDIA kernel module is not loaded or failed to initialize Correct answer.
  4. The GPU is assigned to a MIG instance

Explanation

When the device enumerates on the bus but the driver cannot talk to it, the usual cause is that the kernel module is missing, blacklisted, or failed during load. Checking dmesg and lsmod confirms the module state before deeper hardware diagnosis.

Question 44GPU and Accelerator Configuration

Before enabling MIG on a GPU, what must be true about the device and the environment?

  1. The container runtime must be Docker rather than containerd
  2. The GPU must be running at its maximum application clock
  3. The host must have at least two CPU sockets populated
  4. The GPU model supports MIG and no other processes are using the device when mode is enabled Correct answer.

Explanation

MIG requires hardware support, and switching modes requires the GPU to be idle so the driver can reconfigure the device. After enabling, you create GPU instances and compute instances that match the workload sizes you need.

Question 45GPU and Accelerator Configuration

A two-socket server has GPUs attached to both CPU sockets. A distributed job shows poor throughput because host-to-device copies cross the inter-socket link. What is the best fix?

  1. Lower the GPU power limit to reduce thermal throttling
  2. Pin each rank's host threads to the NUMA node nearest its assigned GPU Correct answer.
  3. Switch the job from NVLink to PCIe peer-to-peer transfers
  4. Disable persistence mode on all GPUs

Explanation

Crossing the inter-socket link adds latency and consumes bandwidth, so aligning host threads with the GPU's local NUMA node keeps copies on the nearest root complex. Tools like nvidia-smi topo -m reveal the correct pairing.

Question 46GPU and Accelerator Configuration

Which signal in nvidia-smi output most directly indicates that a GPU is thermally throttling under sustained load?

  1. A rising ECC corrected error count with stable temperatures
  2. Clocks below the requested application clock while temperature is near the slowdown threshold Correct answer.
  3. The persistence mode field reported as Enabled
  4. The PCIe link width reported as x16

Explanation

Thermal throttling shows up as clocks dropping below the requested target while the die temperature sits at or near the slowdown limit. Correlating the two values confirms that heat, not power capping or workload change, is limiting frequency.

Question 47GPU and Accelerator Configuration

You want to expose a specific GPU to a container by its stable identifier rather than by index. Which value should you put in NVIDIA_VISIBLE_DEVICES?

  1. The CUDA compute capability such as 8.0
  2. The GPU UUID reported by nvidia-smi -L Correct answer.
  3. The driver version string such as 550.54
  4. The PCIe bus address printed by lspci

Explanation

The container toolkit accepts GPU UUIDs, which remain stable across reboots and device reordering. Using a UUID avoids the problem of indices shifting when hardware is added or removed.

Question 48GPU and Accelerator Configuration

After a firmware update, a node boots with only four of eight GPUs visible. Which step best isolates whether the issue is power, seating, or configuration?

  1. Reinstall the CUDA toolkit and rerun the workload
  2. Raise the application clocks on the visible GPUs
  3. Check lspci for the missing devices and review dmesg for PCIe or power faults Correct answer.
  4. Enable MIG on the visible GPUs to recover capacity

Explanation

Comparing lspci enumeration against dmesg messages separates devices that never appeared from ones that appeared but failed to initialize. Power delivery, riser seating, and BIOS settings are the usual causes when GPUs vanish after a firmware change.

Question 49GPU and Accelerator Configuration

Which statement correctly describes how NVLink and PCIe differ for GPU-to-GPU communication in the same node?

  1. NVLink carries only management traffic while PCIe carries tensor data
  2. PCIe provides higher bandwidth than NVLink between GPUs on the same board
  3. NVLink offers higher bidirectional bandwidth and lower latency than PCIe between linked GPUs Correct answer.
  4. NVLink requires a network switch while PCIe connects GPUs directly

Explanation

NVLink is a dedicated high-speed interconnect designed for GPU peer traffic, exceeding PCIe bandwidth and reducing latency for collective operations. PCIe peer-to-peer works but is constrained by the shared switch and root complex.

Question 50GPU and Accelerator Configuration

A user reports that a job fails with an out-of-memory error even though nvidia-smi shows free memory on the GPU. What is the most likely explanation?

  1. The power limit is set below the default value
  2. The GPU has fallen off the bus and is no longer usable
  3. Another process holds a context that reserves memory not shown as used by the reporting process Correct answer.
  4. ECC is disabled on the device, preventing allocation

Explanation

Memory accounting can differ between what a specific process sees and what the device reports, especially when other contexts or MIG instances hold allocations. Checking per-process memory with nvidia-smi and the compute mode clarifies who owns the memory.