NCA-AIIO

Free NCA-AIIO practice questions

10 questions from our NCA-AIIO practice tests, each with the correct answer and a written explanation. Read them with the answers shown, or switch to quiz mode and answer first.

Back to the NCA-AIIO practice exams

Question 1 of 10, Essential AI Knowledge

A team compares hardware needs for training a large language model versus serving it to users. Which statement BEST describes the difference?

  1. ATraining and inference have identical requirements, so the same cluster configuration and numeric precision should be used for both phases of the model lifecycle
  2. BInference needs far more memory than training because it stores gradients and optimizer states, so inference clusters should be sized larger than training clusters
  3. CTraining needs maximum aggregate compute, memory and interconnect bandwidth across many GPUs, whereas inference is driven by latency and throughput per request and often uses lower precision Correct answer
  4. DTraining is latency-bound per request and runs on single GPUs, whereas inference must always span many nodes with high-precision FP64 arithmetic to stay accurate

Answer and explanation: Training processes huge datasets repeatedly and keeps weights, gradients and optimizer states in memory, so it benefits from many GPUs with fast interconnects. Inference serves individual requests where latency and throughput matter, and it commonly uses reduced precision such as FP8 or INT8. Training is not latency-bound per request, the two phases do not have identical needs, and gradients and optimizer states are a training cost rather than an inference cost.

Question 2 of 10, Essential AI Knowledge

An enterprise requires enterprise support, security patching and a validated stack for running AI frameworks and NIM microservices in production on certified servers. Which NVIDIA offering addresses this requirement?

  1. AThe CUDA Toolkit installed on each host
  2. BNVIDIA Omniverse with OpenUSD connectors
  3. CThe free community NGC containers alone
  4. DNVIDIA AI Enterprise software suite Correct answer

Answer and explanation: NVIDIA AI Enterprise is the licensed, supported software platform that includes production branches with security updates, enterprise support and NIM microservices, validated on certified systems and clouds. Community NGC containers are useful but lack the enterprise support commitment. Omniverse targets 3D simulation and digital twins, and the CUDA Toolkit alone is a developer platform rather than a supported end-to-end AI stack.

Question 3 of 10, Essential AI Knowledge

A company wants to deploy a Llama-family LLM on its own Kubernetes cluster quickly, using a prebuilt container that already includes an optimized inference engine and exposes an OpenAI-compatible API. Which NVIDIA offering fits BEST?

  1. ANVIDIA DALI data loading
  2. BNVIDIA RAPIDS cuML library
  3. CNVIDIA NIM microservices Correct answer
  4. DNVIDIA DOCA SDK for DPUs

Answer and explanation: NVIDIA NIM provides prebuilt, optimized inference microservices for popular models with industry-standard APIs, including OpenAI-compatible endpoints for LLMs, so teams can deploy quickly on Kubernetes or a single host. cuML accelerates classical machine learning algorithms, DALI accelerates data loading and augmentation, and DOCA is the SDK for programming BlueField DPUs; none of them packages an LLM serving endpoint.

Question 4 of 10, Essential AI Knowledge

An architect must choose infrastructure for two workloads: weekly fine-tuning of a 70B model and a customer-facing chatbot serving that model. Which statement BEST reflects their different requirements?

  1. AServing favors InfiniBand across hundreds of nodes; fine-tuning needs only one PCIe GPU card
  2. BFine-tuning favors tightly coupled multi-GPU nodes; serving favors latency, scaling and uptime Correct answer
  3. CFine-tuning favors CPUs with large caches; serving favors FP64 GPUs with ECC disabled always
  4. DFine-tuning and serving both need only single small GPUs because the model is already trained

Answer and explanation: Fine-tuning a 70B model still needs substantial GPU memory and fast interconnects, so tightly coupled multi-GPU nodes fit, while serving prioritizes latency, autoscaling and availability. CPUs and FP64 are inappropriate for these workloads, a 70B model does not fit in one small GPU at useful precision, and serving rarely needs a large InfiniBand fabric while fine-tuning needs more than one PCIe card.

Question 5 of 10, AI Infrastructure

What component must be running on an InfiniBand fabric to assign addresses and compute routes before traffic can flow?

  1. AA BGP route reflector placed on each server's BMC interface
  2. BA subnet manager, such as one running on a switch or UFM host Correct answer
  3. CA DHCP server on the management network that leases fabric addresses
  4. DA spanning-tree root bridge elected among the leaf switches

Answer and explanation: InfiniBand fabrics are centrally managed by a subnet manager that discovers the topology, assigns local identifiers (LIDs) and programs forwarding tables; it can run on a managed switch or on a host such as one running Unified Fabric Manager. DHCP leases IP addresses on Ethernet and does not assign InfiniBand LIDs or routes, spanning tree is an Ethernet loop-prevention protocol, and BGP route reflectors on BMCs do not program an InfiniBand fabric.

Question 6 of 10, AI Infrastructure

A company runs a GPU training workload at near 100% utilization every day for the next three years and must keep its proprietary data on premises. Which deployment is MOST cost-effective while meeting the requirement?

  1. ARent on-demand cloud GPU instances and stop them outside business hours only
  2. BUse spot cloud GPU instances because interruptions do not affect long training
  3. CPurchase on-premises GPU systems and amortize the capital cost over the period Correct answer
  4. DUse a public cloud inference API instead of training its own models on premises

Answer and explanation: Consistently high utilization over several years, plus a data residency requirement, favors owning infrastructure, where capital cost amortized across heavy use beats rental pricing. On-demand cloud capacity is cost-effective for bursty or short-term needs and moves the data off premises. Spot instances can be reclaimed, which disrupts training and also breaks the on-premises requirement. A public inference API does not train proprietary models and sends data off site.

Question 7 of 10, AI Infrastructure

What does GPUDirect RDMA enable in a multi-node GPU cluster?

  1. AA network adapter reads and writes GPU memory directly, bypassing host memory copies Correct answer
  2. BTwo GPUs in the same server exchange data over NVLink without involving the driver
  3. CA storage array streams files into GPU memory through the CPU's page cache by default
  4. DThe CPU offloads all of its scheduling decisions to the GPU through a dedicated link

Answer and explanation: GPUDirect RDMA lets an RDMA-capable NIC such as ConnectX access GPU memory directly, removing staging copies through system memory and reducing latency and CPU load for inter-node transfers. GPU-to-GPU transfers within a node use NVLink or peer-to-peer, which is a different feature. Storage-to-GPU direct paths are GPUDirect Storage and specifically avoid the page cache, and GPUs do not take over CPU scheduling.

Question 8 of 10, AI Infrastructure

A team plans an RDMA fabric on standard Ethernet with RoCE v2 for GPU-to-GPU traffic. Which configuration is required to achieve good performance?

  1. ADisable jumbo frames and rely on TCP retransmission for lost packets
  2. BUse spanning tree so only one uplink per leaf carries traffic at a time
  3. CConfigure lossless behavior with PFC and ECN-based congestion control Correct answer
  4. DPlace GPU traffic on the default best-effort class with no QoS policy

Answer and explanation: RoCE v2 performs poorly when packets are dropped, so Ethernet fabrics for RDMA are configured for lossless or near-lossless operation using Priority Flow Control and ECN-driven congestion control (as Spectrum-X does with adaptive routing). RDMA traffic does not rely on TCP retransmission and benefits from jumbo frames. Spanning tree wastes uplink bandwidth, and a best-effort class without QoS risks drops.

Question 9 of 10, AI Operations

A monitoring dashboard shows a training node's GPUs at full utilization but with clocks well below maximum, temperatures near their limit and throttle reasons reporting thermal slowdown. What is the MOST likely cause?

  1. AThe NCCL version is too old, which lowers the GPU clock speed
  2. BInsufficient cooling or airflow is forcing the GPUs to throttle Correct answer
  3. CThe job uses mixed precision, which always halves GPU clocks
  4. DThe scheduler assigned too few CPU cores, which caps GPU clocks

Answer and explanation: When temperatures approach limits and the throttle reason indicates thermal slowdown, the GPU reduces clocks to protect itself, so airflow, cooling capacity or blocked filters should be checked. NCCL versions affect communication, not clocks. Mixed precision does not reduce clock speed, and CPU core allocation can starve a GPU but does not trigger thermal throttling.

Question 10 of 10, AI Operations. Choose 2.

An HPC center runs batch training jobs on a bare-metal GPU cluster and needs fair-share queues, job priorities and backfill scheduling. Which TWO tools are commonly used for this purpose? (Choose two.)

  1. ANVIDIA TensorRT-LLM
  2. BSlurm Workload Manager Correct answer
  3. CNVIDIA Riva
  4. DKubernetes with NVIDIA Run:ai Correct answer
  5. ENVIDIA Nsight Systems

Answer and explanation: Slurm is the dominant HPC batch scheduler with partitions, fair-share, priorities and backfill. Kubernetes paired with NVIDIA Run:ai adds AI-focused scheduling such as quotas, fair-share and GPU fractioning. TensorRT-LLM is an inference optimization library, Nsight Systems is a profiler, and Riva builds speech AI applications; none schedules jobs.