Skip to main content
CertRampLearning

NVIDIA · Free practice questions

NVIDIA NCA-AIIO Practice Questions

15 original NCA-AIIO practice questions spread across NVIDIA's three published exam topics by weight: six on Essential AI Knowledge (38%), six on AI Infrastructure (40%) and three on AI Operations (22%). Each one comes with the answer, a short explanation and the reason every other option is wrong.

Written for the current NVIDIA NCA-AIIO exam · Updated November 11, 2026

These are original practice questions written by CertRamp to the official NVIDIA exam outline. They are not real exam questions and do not come from any exam. Try each question first, then open the answer to see the explanation for every option.

Domain 1

Essential AI Knowledge

Compare and contrast GPU and CPU architectures.FoundationDomain: Essential AI Knowledge

Question 1. A manager asks why the team wants GPUs rather than more CPU servers for deep learning training. Which statement matches how NVIDIA's CUDA documentation describes the two architectures?

  1. Option A: GPUs have much higher single-thread performance than CPUs, so each individual operation finishes sooner
  2. Option B: A GPU runs thousands of threads in parallel, trading single-thread speed for much greater total throughput
  3. Option C: CPUs cannot perform the floating-point matrix operations that neural networks need
  4. Option D: GPUs dedicate more transistors to data caching and flow control, which is why they hide memory latency so well
Show answer and explanation

Correct answer: B. A GPU runs thousands of threads in parallel, trading single-thread speed for much greater total throughput

CPUs are designed for low latency on a serial sequence of operations and run a few tens of threads in parallel. GPUs devote more transistors to data processing and run thousands of threads at once, which suits the highly parallel arithmetic of deep learning.

  • A (incorrect)The opposite is true: a GPU accepts lower single-thread performance in exchange for higher total throughput.
  • B (correct)Correct. The CUDA guide describes the CPU as optimised to run a serial thread as fast as possible and the GPU as optimised for thousands of parallel threads and much greater total throughput.
  • C (incorrect)CPUs can perform these operations; they simply run far fewer of them in parallel.
  • D (incorrect)Dedicating more transistors to caching and flow control is how the CUDA guide characterises CPUs, not GPUs.

Reference: NVIDIA CUDA Programming Guide: Introduction (opens in a new tab)

Compare and contrast training and inference architecture requirements and considerations.FoundationDomain: Essential AI Knowledge

Question 2. Which statement best contrasts the infrastructure considerations for training a large model with those for serving it in production?

  1. Option A: Inference always needs more total compute than training because the model runs continuously
  2. Option B: Training is mainly judged on per-user response time, while inference runs as an offline batch process
  3. Option C: Inference runs only on CPUs, so GPU capacity is needed for training alone
  4. Option D: Training is very compute-intensive; inference balances per-user responsiveness with serving many users
Show answer and explanation

Correct answer: D. Training is very compute-intensive; inference balances per-user responsiveness with serving many users

Training sizing is driven by total compute, data and time to train. Inference sizing is driven by service levels: latency for each request and throughput across concurrent users, which is why inference has its own optimisation techniques.

  • A (incorrect)Training a large model is the heavier workload; NVIDIA cites around 3.8 x 10^25 operations to train Llama 3.1 405B. Inference demand varies with usage.
  • B (incorrect)This reverses the two. Responsiveness to each user is an inference concern; training is a long-running compute job.
  • C (incorrect)Inference is commonly GPU-accelerated, especially for large language models.
  • D (correct)Correct. NVIDIA describes training as enormously compute-hungry and inference as a trade-off between per-user latency and serving many users at once.

Reference: NVIDIA Blog: What's the Difference Between Deep Learning Training and Inference? (opens in a new tab)

Compare and contrast training and inference architecture requirements and considerations.AdvancedDomain: Essential AI Knowledge

Question 3. An LLM serving team sees that the first phase of each request, where all input tokens are processed together, is compute-heavy, while the token-by-token generation that follows behaves differently. Which serving approach runs these two phases on separate GPUs?

  1. Option A: Disaggregated serving, which places the prefill and decode phases on different GPUs
  2. Option B: Speculative decoding, where a smaller model drafts tokens for a larger model to check
  3. Option C: Data-parallel training across several nodes
  4. Option D: Multi-Instance GPU partitioning of a single GPU
Show answer and explanation

Correct answer: A. Disaggregated serving, which places the prefill and decode phases on different GPUs

Prefill processes the whole prompt at once and is computationally heavy; decode generates one token at a time. Because the phases stress hardware differently, disaggregated serving gives prefill its own set of GPUs, separate from the GPUs that run decode.

  • A (correct)Correct. NVIDIA describes disaggregated serving as using separate GPUs for the prefill and decode phases of inference.
  • B (incorrect)Speculative decoding reduces waiting time during generation, but it does not split prefill and decode across GPUs.
  • C (incorrect)Data parallelism is a training technique for spreading batches over GPUs; it is not an inference serving pattern for request phases.
  • D (incorrect)MIG splits one GPU into isolated instances, but it does not by itself assign the prefill and decode phases to different hardware.

Reference: NVIDIA Blog: What's the Difference Between Deep Learning Training and Inference? (opens in a new tab)

Describe the NVIDIA software stack used in an AI environment.FoundationDomain: Essential AI Knowledge

Question 4. A platform team wants a trusted source for GPU-optimised framework containers, pretrained models and Kubernetes Helm charts. Which part of the NVIDIA software stack provides this?

  1. Option A: Slurm, the workload manager that queues, schedules and launches multi-node batch jobs
  2. Option B: The baseboard management controller (BMC) built into each DGX system
  3. Option C: The NGC catalog of GPU-optimised containers, pretrained models and Helm charts
  4. Option D: The NVIDIA License System that serves licenses to vGPU client VMs
Show answer and explanation

Correct answer: C. The NGC catalog of GPU-optimised containers, pretrained models and Helm charts

NGC containers bundle frameworks, libraries and dependencies so they run consistently from cloud to data centre to edge. Pretrained models can be used for inference or fine-tuning, and Helm charts help deploy GPU-optimised applications on Kubernetes.

  • A (incorrect)Slurm is a workload manager for scheduling jobs; it does not distribute containers or models.
  • B (incorrect)The BMC is for out-of-band hardware management, not for publishing software.
  • C (correct)Correct. NVIDIA describes the NGC catalog as curated GPU-optimised software for AI, HPC and visualisation, with every image security scanned and performance tested on GPUs.
  • D (incorrect)The NVIDIA License System licenses vGPU software; it is not a software catalogue.

Reference: NVIDIA NGC Catalog User Guide (opens in a new tab)

Describe the software components related to the life cycle of AI development and deployment.IntermediateDomain: Essential AI Knowledge

Question 5. A company has trained models in PyTorch, exported others to ONNX and optimised some with TensorRT. It wants one inference server that exposes HTTP/REST and gRPC endpoints, batches requests dynamically and reports latency and GPU utilisation metrics. Which NVIDIA component fits?

  1. Option A: NVIDIA Data Center GPU Manager (DCGM)
  2. Option B: NVIDIA Unified Fabric Manager (UFM)
  3. Option C: NVIDIA Triton Inference Server
  4. Option D: NVIDIA Base Command Manager
Show answer and explanation

Correct answer: C. NVIDIA Triton Inference Server

In the deployment stage of the AI life cycle, an inference server turns trained models into network services. Triton's multi-framework backends, dynamic batching and built-in metrics make it NVIDIA's general-purpose answer for that stage.

  • A (incorrect)DCGM monitors and diagnoses GPUs; it does not serve models.
  • B (incorrect)UFM manages InfiniBand fabrics and their telemetry, not inference.
  • C (correct)Correct. Triton deploys models from TensorRT, PyTorch, ONNX and other frameworks, supports dynamic batching and concurrent model execution, and exposes HTTP/REST and gRPC protocols with metrics.
  • D (incorrect)Base Command Manager provisions and administers clusters; it is not a model server.

Reference: NVIDIA Triton Inference Server documentation (opens in a new tab)

Explain the purpose and use case of various NVIDIA solutions.IntermediateDomain: Essential AI Knowledge

Question 6. In NVIDIA's DGX SuperPOD reference architecture, which component is described as a software suite that gives developers AI tools and frameworks and gives IT teams orchestration and management tools?

  1. Option A: NVIDIA AI Enterprise
  2. Option B: NVIDIA Unified Fabric Manager
  3. Option C: Slurm
  4. Option D: NVIDIA Base Command Manager
Show answer and explanation

Correct answer: A. NVIDIA AI Enterprise

A SuperPOD combines several NVIDIA solutions with distinct jobs: Base Command Manager runs the cluster, UFM runs the fabric, Slurm schedules jobs and NVIDIA AI Enterprise supplies the AI software. Knowing which product does what is a recurring theme of this objective.

  • A (correct)Correct. The reference architecture describes NVIDIA AI Enterprise as a suite of AI development tools and frameworks plus orchestration and management tools for IT.
  • B (incorrect)UFM manages the InfiniBand compute and storage fabrics with real-time telemetry; it is not a developer software suite.
  • C (incorrect)Slurm is the workload scheduler for multi-node batch jobs, not a suite of AI frameworks.
  • D (incorrect)Base Command Manager is the cluster management solution for provisioning and administration, not the AI framework suite.

Reference: NVIDIA DGX SuperPOD Reference Architecture (DGX H100): Key Components (opens in a new tab)

Domain 2

AI Infrastructure

Explain the purpose and benefits of a DPU in a data center.FoundationDomain: AI Infrastructure

Question 7. An architect proposes adding NVIDIA BlueField DPUs to GPU servers. What is the main benefit she is after?

  1. Option A: Additional Tensor Cores on the server that speed up model training and fine-tuning
  2. Option B: Replacing the GPUs for inference workloads at a lower cost per served request
  3. Option C: High-bandwidth GPU-to-GPU links between the GPUs inside a single server
  4. Option D: Offloading networking, storage, security and management work from the host CPU
Show answer and explanation

Correct answer: D. Offloading networking, storage, security and management work from the host CPU

Moving infrastructure services onto the DPU frees host CPU cycles for applications and separates infrastructure from business workloads, which supports a zero-trust design. BlueField is programmable through the NVIDIA DOCA framework.

  • A (incorrect)A DPU does not add GPU compute; training throughput still comes from the GPUs.
  • B (incorrect)A DPU is an infrastructure processor, not a substitute for GPUs running models.
  • C (incorrect)GPU-to-GPU connectivity inside a DGX system comes from NVLink and NVSwitch, not a DPU.
  • D (correct)Correct. NVIDIA describes BlueField-3 as offloading, accelerating and isolating software-defined networking, storage, security and management, running its own software on Arm cores.

Reference: NVIDIA BlueField-3 Networking Platform User Guide: Introduction (opens in a new tab)

Determine networking requirements for AI workloads.IntermediateDomain: AI Infrastructure

Question 8. In a multi-node training job, profiling shows that gradients are copied from GPU memory into host memory before the network adapter sends them. Which NVIDIA technology provides a direct data path between the GPU and the network adapter?

  1. Option A: Multi-Instance GPU (MIG)
  2. Option B: GPUDirect RDMA
  3. Option C: Time-sliced vGPU
  4. Option D: NVLink between the GPUs inside the server
Show answer and explanation

Correct answer: B. GPUDirect RDMA

Avoiding the copy to host memory reduces latency and CPU overhead for inter-node communication. Distributed training depends heavily on this kind of fast GPU-to-network path in addition to a high-bandwidth fabric.

  • A (incorrect)MIG partitions a GPU into isolated instances; it does not change how data moves to the network adapter.
  • B (correct)Correct. GPUDirect RDMA enables a direct path for data exchange between the GPU and a third-party peer device such as a network interface, using standard PCI Express features.
  • C (incorrect)Time-sliced vGPU shares a GPU between virtual machines; it is unrelated to the network data path.
  • D (incorrect)NVLink connects GPUs to each other within a system; it does not give the network adapter direct access to GPU memory.

Reference: NVIDIA GPUDirect RDMA documentation (opens in a new tab)

Identify high-speed data center network options and their use cases.IntermediateDomain: AI Infrastructure

Question 9. In the DGX SuperPOD reference architecture for DGX H100, which network carries the GPU traffic between nodes during a multi-node training job?

  1. Option A: The rail-aligned InfiniBand compute fabric
  2. Option B: The in-band management Ethernet network
  3. Option C: The out-of-band management Ethernet network
  4. Option D: NVLink and NVSwitch inside each DGX system
Show answer and explanation

Correct answer: A. The rail-aligned InfiniBand compute fabric

A SuperPOD separates traffic into a compute fabric, a storage fabric, an in-band management network and an out-of-band management network. Inside a node, GPUs talk over NVLink; between nodes, they talk over the InfiniBand compute fabric.

  • A (correct)Correct. The compute fabric connects the GPU nodes over InfiniBand; traffic between nodes or between rails traverses the spine layer.
  • B (incorrect)The in-band Ethernet network serves cluster management services, home file systems and access to external services such as the NGC registry.
  • C (incorrect)The out-of-band network connects management ports of systems, switches and PDUs; it does not carry workload traffic.
  • D (incorrect)NVLink and NVSwitch connect the eight GPUs within one DGX H100. Traffic between nodes needs the compute fabric.

Reference: NVIDIA DGX SuperPOD Reference Architecture (DGX H100): Network Fabrics (opens in a new tab)

Identify key components and considerations of a cluster of an accelerated infrastructure.FoundationDomain: AI Infrastructure

Question 10. A cluster design needs a network that connects the management ports of the DGX systems, management servers, storage, network switches and rack PDUs. Which SuperPOD network serves this role?

  1. Option A: The compute fabric
  2. Option B: The storage fabric
  3. Option C: The in-band management network
  4. Option D: The out-of-band management network
Show answer and explanation

Correct answer: D. The out-of-band management network

Keeping device management on a separate out-of-band network means administrators can still reach hardware when the data networks or operating systems are down. It is a standard building block of an accelerated cluster.

  • A (incorrect)The compute fabric carries GPU traffic between nodes, not device management ports.
  • B (incorrect)The storage fabric carries storage I/O to the high-performance file system.
  • C (incorrect)The in-band network carries management services and user-facing access such as home directories and NGC; it is not the network for device management ports.
  • D (correct)Correct. The reference architecture's out-of-band Ethernet fabric connects the management ports of DGX systems, management servers, storage, networking gear and rack PDUs.

Reference: NVIDIA DGX SuperPOD Reference Architecture (DGX H100): Network Fabrics (opens in a new tab)

Scale a GPU infrastructure for different use cases.AdvancedDomain: AI Infrastructure

Question 11. Several small inference services each use only a fraction of an H100 GPU. The owners need predictable latency and want a fault in one service not to affect the others. What is the most suitable approach?

  1. Option A: Give each service its own full GPU so that nothing is shared
  2. Option B: Run all services as ordinary processes on one GPU without partitioning
  3. Option C: Partition the GPU into isolated Multi-Instance GPU (MIG) instances
  4. Option D: Add a BlueField DPU to the server
Show answer and explanation

Correct answer: C. Partition the GPU into isolated Multi-Instance GPU (MIG) instances

MIG assigns L2 cache banks, memory controllers and SMs to each instance, so one tenant cannot affect another's throughput or scheduling. That makes it a good way to scale right-sized GPU capacity for many small workloads.

  • A (incorrect)This isolates the services but leaves most of each GPU idle, which is the utilisation problem MIG is designed to solve.
  • B (incorrect)Unpartitioned sharing lets one workload thrash caches or saturate memory bandwidth for the others, so latency is not predictable.
  • C (correct)Correct. MIG splits supported GPUs into up to seven instances with separate paths through the memory system, a defined quality of service and fault isolation.
  • D (incorrect)A DPU offloads infrastructure functions; it does not partition GPU compute or memory between workloads.

Reference: NVIDIA Multi-Instance GPU User Guide: Introduction (opens in a new tab)

Identify key concepts and high-level specifications related to power and cooling requirements within a data center.IntermediateDomain: AI Infrastructure

Question 12. The DGX H100 user guide lists six 3.3 kW power supplies in a 4+2 redundant configuration. What does the 4+2 configuration mean?

  1. Option A: All six supplies must be installed and working before the system is able to power on
  2. Option B: Four supplies carry the load and up to two can fail without stopping the system
  3. Option C: The complete system draws at most 3.3 kW at full load
  4. Option D: Four supplies power the GPUs and two supplies power the CPUs and motherboard
Show answer and explanation

Correct answer: B. Four supplies carry the load and up to two can fail without stopping the system

Power redundancy notation tells facility planners how many supplies are needed and how many can fail. With the DGX H100 user guide listing a maximum of 10.2 kW for a single system, rack power and cooling must be planned for far higher density than typical CPU servers.

  • A (incorrect)A redundant configuration exists precisely so that the system does not depend on every supply.
  • B (correct)Correct. In a 4+2 design four supplies are sufficient for operation and the remaining two provide redundancy.
  • C (incorrect)3.3 kW is the rating of each supply; the system is specified for considerably more, up to 10.2 kW.
  • D (incorrect)The notation describes redundancy, not a split of supplies between components.

Reference: NVIDIA DGX H100 User Guide: Introduction to the DGX H100 System (opens in a new tab)

Domain 3

AI Operations

Describe AI data center management and monitoring essentials.IntermediateDomain: AI Operations

Question 13. After a GPU board is replaced, an administrator wants the most thorough DCGM diagnostic, including memory, PCIe/NVLink and stress tests, before returning the node to production. Which command fits?

  1. Option A: dcgmi diag -r 1
  2. Option B: dcgmi health -c
  3. Option C: dcgmi diag -r 2
  4. Option D: dcgmi diag -r 3
Show answer and explanation

Correct answer: D. dcgmi diag -r 3

DCGM separates passive health monitoring, which runs alongside jobs, from active diagnostics, which need the GPU to themselves. Deeper diagnostic levels take longer and are the right choice after hardware changes.

  • A (incorrect)Level 1 is a quick check that runs in seconds; it does not include stress testing.
  • B (incorrect)This reports the status of the background health watches. It is non-invasive and does not run active stress tests.
  • C (incorrect)Level 2 is a medium run of about two minutes; it is less thorough than level 3.
  • D (correct)Correct. Level 3 is the long diagnostic of about 15 minutes and includes memory, PCIe/NVLink and stress tests.

Reference: NVIDIA DCGM User Guide: Feature Overview (opens in a new tab)

Describe AI cluster orchestration and job scheduling essentials.FoundationDomain: AI Operations

Question 14. A Kubernetes team wants to automate installing and managing the NVIDIA driver, the container toolkit, the Kubernetes device plugin, GPU Feature Discovery and DCGM-based monitoring on every GPU node. What should it deploy?

  1. Option A: The NVIDIA GPU Operator
  2. Option B: Slurm
  3. Option C: NVIDIA Unified Fabric Manager
  4. Option D: The NVIDIA Virtual GPU Manager
Show answer and explanation

Correct answer: A. The NVIDIA GPU Operator

Without the operator, each GPU node needs drivers, runtime, device plugin and monitoring installed and kept in step by hand. The GPU Operator handles that life cycle so Kubernetes can schedule GPU workloads reliably.

  • A (correct)Correct. The GPU Operator uses the Kubernetes operator framework to automate the management of all NVIDIA software components needed to provision GPUs.
  • B (incorrect)Slurm schedules batch jobs; it does not install and manage NVIDIA software on Kubernetes nodes.
  • C (incorrect)UFM manages InfiniBand fabrics, not GPU node software in Kubernetes.
  • D (incorrect)The Virtual GPU Manager runs in a hypervisor to share GPUs between virtual machines; it does not manage Kubernetes node components.

Reference: NVIDIA GPU Operator: Overview (opens in a new tab)

Identify the key considerations for virtualizing accelerated infrastructure.IntermediateDomain: AI Operations

Question 15. A virtual machine running a large training job must have exclusive use of an entire physical GPU, with no sharing with other VMs. Which virtualization option fits?

  1. Option A: A time-sliced vGPU
  2. Option B: A MIG-backed vGPU
  3. Option C: GPU passthrough
  4. Option D: Installing the NVIDIA License System client in the VM
Show answer and explanation

Correct answer: C. GPU passthrough

NVIDIA vGPU lets multiple VMs share a GPU, either time-sliced or backed by MIG instances. Passthrough is the choice when one VM needs the whole GPU and sharing is not wanted.

  • A (incorrect)Time-sliced vGPU lets several VMs share one physical GPU, which is not exclusive use.
  • B (incorrect)A MIG-backed vGPU maps to one GPU instance, a partition of the physical GPU, rather than the whole device.
  • C (correct)Correct. With passthrough the entire physical GPU is assigned to one VM, bypassing the NVIDIA Virtual GPU Manager, and is used only by that VM's driver.
  • D (incorrect)Licensing is a separate requirement for vGPU software; it does not decide how the GPU is allocated.

Reference: NVIDIA vGPU Software User Guide: Introduction (opens in a new tab)

Exam information last verified: November 13, 2026

Exam details can change. Always confirm them with the official provider before you book.

Want more practice?

15 questions are a sample. The exam is 50 questions items.

The full CertRamp NVIDIA NCA-AIIO practice exams have 6 tests that grow with you, from a diagnostic start to an exam-level Test 4 and a challenge set beyond the exam, with an explanation for every option.

Who writes the exams

Written by a practitioner, not a question mill

Joshua Ravnjak, founder of CertRamp Learning

Joshua Ravnjak

Senior Consultant, digital transformation and AI

Founder, CertRamp Learning

921 learners and 61 reviews on Udemy (as of November 3, 2026)

I have more than 12 years of professional experience and work as a senior consultant in an IT management consultancy, focusing on digital transformation, AI and requirements engineering in the insurance industry. I have earned many certifications myself, so I know what exam preparation feels like. That is why every CertRamp practice exam follows the same progression: an easy start to see where you stand, then real exam level, then beyond it. You walk into the exam with a buffer and a good feeling.

Selected certifications

  • IREB® CPRE Foundation Level
  • ITIL® 4 Foundation
  • Professional Scrum Master I (PSM I)
  • Professional Scrum Product Owner I (PSPO I)
  • Certified SAFe® 6 Scrum Master
  • Certified AI-Empowered SAFe® 6 Product Owner/Product Manager

Plus further certifications from Microsoft and Anthropic (Claude).

ITIL® is a registered trademark of PeopleCert. IREB® is a registered trademark of IREB e.V. SAFe® is a registered trademark of Scaled Agile, Inc. PSM and PSPO are trademarks of Scrum.org.

Also in print

Cover: CPRE Foundation Level Prüfungsvorbereitung 2026 by Joshua Ravnjak

CPRE Foundation Level Prüfungsvorbereitung 2026

6 Übungsprüfungen mit 270 Fragen und ausführlichen Erklärungen

Paperback, German, October 2026

My CPRE Foundation Level exam-prep book (in German) follows the same progressive approach.

View on Amazon (opens in a new tab)