Question bank
NVIDIA Certified Associate - AI Infrastructure
41 NVIDIA Certified Associate - AI Infrastructure questions with the correct answer and a written explanation, organised across 3 exam domains.
- 41 questions
- 3 domains
- Foundation level
certilumQuestion index
Jump straight to any question. Answers and explanations are inline.
AI Infrastructure
16 questions in this domain
- Question 1Single Choice
What is the primary purpose of NVLink in an AI infrastructure cluster?
- AInter-node cluster management protocol
- BHigh-bandwidth, low-latency intra-node GPU interconnectCorrect answer
- CNetwork-attached storage protocol
- DGPU virtualization hypervisor
Why this is the answer
NVLink is a high-bandwidth, low-latency interconnect for GPU-to-GPU communication within a node, not for management, storage, or virtualization.
- Question 2Multiple Select
Which fabric types are recognized for their RDMA capabilities in AI clusters?
- ANVLink
- BInfiniBandCorrect answer
- CEthernetCorrect answer
- DPCIe
Why this is the answer
Both InfiniBand and Ethernet can support RDMA, while NVLink is a GPU interconnect and PCIe is a local bus, not a network fabric.
- Question 3Interactive
An engineer issues the command shown to inspect GPU health. What class of error is being queried?
- ACRC errors on InfiniBand links
- BECC memory errorsCorrect answer
- CPCIe training errors
- DGPU thermal throttling
Why this is the answer
The query reads corrected ECC error counters, so it monitors GPU memory reliability rather than fabrics or thermal state.
- Question 4Single Choice
What is the function of the NVIDIA Collective Communications Library (NCCL)?
- AIt orchestrates communication between GPUsCorrect answer
- BIt manages Kubernetes resource scheduling
- CIt performs storage replication
- DIt provides GPU power management
Why this is the answer
NCCL is the software layer that coordinates GPU communication, and its diagnostics focus on timeouts and ring/tree topologies.
- Question 5Multiple Select
Which of the following are documented common traps when troubleshooting AI infrastructure?
- AConfusing GPU utilization with GPU throughputCorrect answer
- BBlaming the algorithm when the root cause is an NCCL topologyCorrect answer
- CIgnoring the physical layer, such as unseated cablesCorrect answer
- DCalling GPUDirect RDMA a CPU-based transfer mechanism
Why this is the answer
The playbook lists utilization/throughput confusion, blaming the algorithm, and ignoring physical-layer issues as common traps.
- Question 6Interactive
The output shown is from an InfiniBand link diagnostic. What type of problem does this status indicate?
- ALink-state errorCorrect answer
- BEthernet congestion-induced latency
- CNCCL timeout caused by environment variables
- DGPU ECC error
Why this is the answer
A Down state with a disabled physical state corresponds to a link-state error in the InfiniBand fabric, not congestion or memory errors.
- Question 7Single Choice
What is a primary benefit of GPUDirect RDMA in AI clusters?
- AIt increases GPU core clock speeds
- BIt allows data to move between GPU memory and network adapters without CPU involvementCorrect answer
- CIt replaces NVLink with Ethernet
- DIt eliminates the need for InfiniBand subnet managers
Why this is the answer
GPUDirect RDMA enables direct transfers between GPU memory and network adapters, bypassing the CPU to reduce latency.
- Question 8Multiple Select
Which indicators are associated with physical-layer link problems in AI infrastructure?
- AECC errors on GPU memoryCorrect answer
- BCRC errors on InfiniBand linksCorrect answer
- CPCIe training errorsCorrect answer
- DHigher application-level checkpoint frequency
Why this is the answer
The playbook lists ECC errors, InfiniBand CRC errors, and PCIe training errors as signs of faulty cabling, loose connections, or failing transceivers.
- Question 9Interactive
An administrator runs the command shown to inspect GPU communication paths. Which interconnect is being validated?
- AInfiniBand fabric routing
- BNVLink/NVSwitch topologyCorrect answer
- CEthernet VLAN configuration
- DPCIe power management
Why this is the answer
The nvidia-smi topology command displays GPU connectivity, including NVLink links and switches, to detect broken or incorrectly detected topologies.
- Question 10Single Choice
What does the trap 'confusing throughput with latency' describe?
- AAssuming lower bandwidth always reduces accuracy
- BReplacing low latency with high throughput
- CAssuming a high-bandwidth link is healthy even when latency is highCorrect answer
- DUsing throughput to estimate GPU utilization
Why this is the answer
High bandwidth can mask high latency, and in AI training, latency can be more damaging than reduced throughput.
- Question 11Multiple Select
Which statements about NVLink are correct?
- AIt is used as a long-haul inter-node fabric
- BIt is intended for intra-node GPU communicationCorrect answer
- CIt is based on IEEE Ethernet standards
- DIt is a high-bandwidth, low-latency interconnectCorrect answer
Why this is the answer
NVLink is an intra-node GPU interconnect with high bandwidth and low latency; NVSwitch scales it within the node.
- Question 12Interactive
The environment variables shown are being set before launching a distributed job. Which software layer are they configuring?
- ANCCL communication between GPUsCorrect answer
- BInfiniBand subnet manager behavior
- CKubernetes pod scheduling
- DGPU thermal limits
Why this is the answer
These variables control NCCL debug output and peer-to-peer level, so they affect NCCL communication and topology detection.
- Question 13Single Choice
Why is MIG not intended for large-scale distributed training?
- AIt is a thermal throttling mechanism
- BIt replaces NVLink for multi-node communication
- CMIG is designed to partition a GPU, not to pool GPUs for a single distributed jobCorrect answer
- DIt automatically load balances across Ethernet fabrics
Why this is the answer
MIG partitions a physical GPU into isolated instances; it is not meant to create a large distributed training pool.
- Question 14Multiple Select
Which steps are appropriate when troubleshooting GPUDirect RDMA issues?
- AVerify PCIe peer-to-peer supportCorrect answer
- BDisable all ECC error reporting
- CAlways use NVLink to connect GPUs across nodes
- DConfirm correct kernel drivers for GPUDirect RDMACorrect answer
Why this is the answer
GPUDirect RDMA depends on PCIe peer-to-peer support and kernel drivers that allow direct GPU-to-adapter transfers.
- Question 15Interactive
The output of the command shown indicates a lower clock speed while the temperature is high. What is the most likely root cause?
- ABroken NVLink link
- BThermal throttlingCorrect answer
- CInfiniBand subnet manager failure
- DNCCL topology mismatch
Why this is the answer
High temperature causes the GPU to reduce clock speed, a condition known as thermal throttling.
- Question 16Single Choice
Why does horizontal scaling introduce extra communication overhead?
- AAdditional network communication is required, which can reduce scaling efficiencyCorrect answer
- BLatency disappears because NVLink connects all nodes
- CGPU memory automatically becomes unlimited
- DNCCL no longer needs to coordinate
Why this is the answer
Adding nodes requires data exchange over the network, and communication overhead can limit scaling benefits.
AI Operations
9 questions in this domain
- Question 17Single Choice
What is the primary purpose of NVLink and NVSwitch in an AI infrastructure?
- AProvide high-bandwidth, low-latency intra-node GPU communicationCorrect answer
- BConnect GPUs to remote storage arrays
- CManage container orchestration
- DReplace Ethernet for inter-node traffic
Why this is the answer
NVLink and NVSwitch enable high-bandwidth, low-latency interconnect for intra-node GPU communication, supporting optimal peer-to-peer memory access.
- Question 18Multiple Select
Which two of the following are typical InfiniBand fabric troubleshooting areas?
- ALink-state errorsCorrect answer
- BSubnet manager issuesCorrect answer
- CVLAN misconfiguration
- DEthernet congestion collapse
Why this is the answer
Link-state errors and subnet manager issues are common InfiniBand concerns. VLAN and Ethernet congestion are associated with Ethernet environments, not InfiniBand.
- Question 19Interactive
Review the displayed command and identify what it is used to troubleshoot.
- ANVLink topology detectionCorrect answer
- BEthernet packet loss measurement
- CKernel module loading
- DMemory bandwidth benchmarking
Why this is the answer
The command displays GPU peer-to-peer topology, which is used to detect NVLink connections and verify correct topology detection.
- Question 20Single Choice
What capability does GPUDirect RDMA primarily enable?
- ADirect data transfer between GPU memory and network adaptersCorrect answer
- BCPU-side encryption of network traffic
- CVirtual GPU memory overcommit
- DAutomatic GPU driver updates
Why this is the answer
GPUDirect RDMA allows GPUs to exchange data directly with network adapters, bypassing the CPU and reducing latency.
- Question 21Multiple Select
Which error detection mechanisms can identify faulty cabling or failing transceivers?
- AECC errorsCorrect answer
- BCRC errors on InfiniBand linksCorrect answer
- CPCIe training errorsCorrect answer
- DHypervisor migration logs
Why this is the answer
ECC, CRC, and PCIe training errors are physical-layer indicators. Hypervisor migration logs are not used to detect physical-layer failures.
- Question 22Interactive
The given command reports GPU interconnect status. Which tool is being used?
- Anvidia-smi nvlink -sCorrect answer
- Bibstatus
- Cnvtop
- Dnvcc
Why this is the answer
The command 'nvidia-smi nvlink -s' queries NVLink status and is part of the NVIDIA System Management Interface.
- Question 23Single Choice
Which trap is described by assuming a high-bandwidth link works well despite high latency?
- AConfusing Throughput with LatencyCorrect answer
- BBlaming the Algorithm
- CIgnoring the Physical Layer
- DMisinterpreting MIG
Why this is the answer
High latency can be more damaging than lower throughput in AI training. This trap confuses throughput with latency.
- Question 24Multiple Select
Which of the following can cause an inefficient NCCL ring or tree topology?
- AMisconfigured environment variablesCorrect answer
- BNetwork mismatchesCorrect answer
- CHigher GPU memory capacity
- DLarger model parameter count
Why this is the answer
NCCL errors and topology inefficiencies often stem from misconfigured environment variables or network mismatches, not from model size or memory capacity.
- Question 25Interactive
What performance metric is the displayed command designed to test?
- AAll-reduce bandwidth and latencyCorrect answer
- BGPU thermal throttling levels
- CDisk write throughput
- DMemory ECC error rate
Why this is the answer
The command runs an all-reduce performance benchmark, measuring collective communication bandwidth and latency across processes.
Essential AI Knowledge
16 questions in this domain
- Question 26Single Choice
Which statement best describes the primary purpose of NVLink and NVSwitch in an AI infrastructure?
- AEnabling communication between GPUs inside one serverCorrect answer
- BConnecting GPUs across separate data centers
- CConnecting storage devices to the network
- DScheduling AI workload containers
Why this is the answer
NVLink and NVSwitch are high-bandwidth, low-latency interconnects designed for intra-node GPU communication, not inter-node or storage networking.
- Question 27Multiple Select
Which statements correctly describe architectural differences between InfiniBand and Ethernet fabrics for RDMA? Select all that apply.
- AInfiniBand uses a subnet manager for fabric control.Correct answer
- BEthernet can experience congestion-induced latency.Correct answer
- CEthernet is inherently lossless for RDMA.
- DInfiniBand depends on TCP/IP for flow control.
Why this is the answer
InfiniBand fabrics require a subnet manager, and Ethernet RDMA can suffer from congestion-induced latency. TCP/IP is not part of InfiniBand flow control.
- Question 28Interactive
Examine the initialization log in the code block. Which software component is being initialized?
- ANCCLCorrect answer
- BNVSwitch
- CGPUDirect RDMA
- DEthernet fabric
Why this is the answer
The log shows NCCL version and initialization messages, indicating the NVIDIA Collective Communications Library is the software being initialized.
- Question 29Single Choice
What does GPUDirect RDMA enable in high-performance AI clusters?
- ADirect data transfer between GPU memory and network adaptersCorrect answer
- BReplacing network adapters with NVLink switches
- CIncreasing GPU core clock speed
- DManaging Kubernetes pods automatically
Why this is the answer
GPUDirect RDMA allows direct data transfer between GPU memory and network adapters, reducing latency by bypassing the CPU.
- Question 30Multiple Select
Which of the following are physical-layer error detection mechanisms mentioned in the playbook? Select all that apply.
- AMonitoring ECC errorsCorrect answer
- BChecking CRC errors on InfiniBand linksCorrect answer
- CDetecting PCIe training errorsCorrect answer
- DMeasuring GPU clock speed only
Why this is the answer
The playbook lists ECC errors, CRC errors on InfiniBand links, and PCIe training errors as physical-layer detection mechanisms.
- Question 31Interactive
The code block shows output listing multiple GPU links. Which interconnect is being checked?
- ANVLinkCorrect answer
- BInfiniBand
- CEthernet
- DPCIe
Why this is the answer
The per-GPU link bandwidth output is typical of NVLink status checks, such as nvidia-smi nvlink -s.
- Question 32Single Choice
According to the playbook, what is a primary diagnostic focus for NCCL?
- AResolving NCCL errors like timeouts or incorrect ring/tree topologiesCorrect answer
- BAdjusting GPU voltage and fan curves
- CReplacing NVLink cables
- DResizing Kubernetes nodes
Why this is the answer
NCCL troubleshooting focuses on resolving errors such as timeouts and incorrect ring/tree topologies caused by configuration or network mismatches.
- Question 33Multiple Select
Which statements correctly identify intra-node and inter-node interconnect technologies? Select all that apply.
- ANVLink is used for intra-node GPU communication.Correct answer
- BNVLink is an inter-node fabric technology.
- CInfiniBand is commonly used for inter-node fabric communication.Correct answer
- DInfiniBand is used inside a single GPU package.
Why this is the answer
NVLink connects GPUs within one node, while InfiniBand is an inter-node fabric used for RDMA-based communication between servers.
- Question 34Interactive
The code block contains a connection log line. Which fabric is NCCL using?
- AInfiniBandCorrect answer
- BEthernet
- CNVLink
- DPCIe
Why this is the answer
NET/IB in the NCCL log indicates the InfiniBand fabric is being used for communication.
- Question 35Single Choice
Why can a high-bandwidth link still harm AI training performance, according to the playbook?
- AHigh latency can be more damaging than lower throughputCorrect answer
- BBandwidth never affects distributed training
- CHigh bandwidth always means low latency
- DLatency is irrelevant to collective operations
Why this is the answer
The playbook warns that confusing throughput with latency is a trap: high latency can hurt AI training more than lower bandwidth.
- Question 36Multiple Select
Which actions are part of troubleshooting physical-layer and peer-to-peer connectivity? Select all that apply.
- AIdentifying broken NVLink linksCorrect answer
- BMonitoring ECC errorsCorrect answer
- CVerifying PCIe peer-to-peer supportCorrect answer
- DBlaming the model architecture
Why this is the answer
Broken NVLink links, ECC errors, and PCIe peer-to-peer verification are physical-layer or connectivity checks. Blaming the algorithm is a common trap.
- Question 37Interactive
An engineer sees the GPU status in the code block and says the GPU is fully busy. Why might this be misleading?
- AGPU utilization is not the same as actual computational throughputCorrect answer
- BMemory usage is too low to compute
- CClock speed is always constant
- DUtilization can never be near 100%
Why this is the answer
High utilization does not guarantee high throughput; confusing utilization with actual work done is a common diagnostic trap.
- Question 38Single Choice
What is the NVIDIA Collective Communications Library (NCCL) in an AI cluster?
- ASoftware layer that orchestrates communication between GPUsCorrect answer
- BA physical switch for GPU interconnect
- CA driver for InfiniBand network cards
- DA cloud orchestration platform
Why this is the answer
NCCL is the software layer that coordinates communication between GPUs, such as ring and tree collectives.
- Question 39Multiple Select
Which statements represent common traps described in the playbook? Select all that apply.
- AAttributing performance drops to the algorithm when NCCL topology is inefficientCorrect answer
- BOverlooking simple hardware issues like unseated cablesCorrect answer
- CAssuming high bandwidth alone means the link is healthyCorrect answer
- DUsing NVSwitch for storage traffic is a best practice
Why this is the answer
The playbook lists blaming the algorithm, ignoring the physical layer, and confusing throughput with latency as common traps.
- Question 40Interactive
The code block shows MIG instances on a GPU. Which statement reflects a common misinterpretation of MIG?
- AMIG is used for large-scale distributed trainingCorrect answer
- BMIG allows partitioning a GPU into smaller compute instances
- CMIG improves NVLink bandwidth
- DMIG is a network protocol for InfiniBand
Why this is the answer
The playbook warns that MIG is often mistaken for large-scale distributed training, while it is actually meant to partition GPUs into smaller instances.
- Question 41Single Choice
Which of the following is a recommended troubleshooting step for GPUDirect RDMA?
- AVerifying PCIe peer-to-peer support and correct kernel driversCorrect answer
- BReplacing GPU memory modules
- CIncreasing the number of Ethernet switches
- DDisabling error correction on all links
Why this is the answer
GPUDirect RDMA troubleshooting involves verifying PCIe peer-to-peer support, kernel drivers, and topology to enable CPU bypass.
Frequently asked questions
Everything about the NVIDIA Certified Associate - AI Infrastructure question bank and how to use it.