Overview
Domain 1 builds the conceptual base for the rest of the NCA-AIIO exam. Its eight objectives cover the NVIDIA software stack, the different architecture needs of training and inference, the hierarchy of AI, machine learning and deep learning, the drivers behind recent AI adoption, major use cases and industries, the purpose of NVIDIA solutions such as DGX, NVIDIA AI Enterprise and NIM microservices, the software used across the AI development and deployment lifecycle, and the contrast between GPU and CPU designs.
At 38%, this is the second-largest domain, close behind AI Infrastructure. It is associate level, so questions ask you to recognize which concept, product or architecture fits a short scenario rather than to configure anything. The vocabulary learned here, such as Tensor Cores, CUDA, TensorRT and Triton, reappears in the infrastructure and operations domains.
How it's tested
NCA-AIIO questions in this domain are short recognition items. A scenario names a workload, a business goal or a lifecycle stage, and you pick the concept, NVIDIA product or architecture that fits. Expect product-to-purpose matching, definitional contrasts and questions about why GPUs accelerate AI workloads.
-
Match the NVIDIA product
A need such as optimizing a trained model, serving it at scale or packaging it as a container is described, and you pick TensorRT, Triton, NIM or NGC.
-
Place it in the hierarchy
A technique such as a neural network or a rules engine is described, and you decide whether it counts as AI, machine learning or deep learning.
-
Training or inference
A workload profile is described, and you identify whether it is training or inference and what that implies for compute, memory and precision.
A team has a trained PyTorch model and needs lower latency in production on NVIDIA GPUs. Retraining with more data does not address latency, and Slurm is a scheduler. The fit is TensorRT, which optimizes a trained network with layer fusion and reduced precision such as FP8 or INT8. Spotting that the model is already trained and the goal is faster inference points to an inference optimizer rather than a training tool.
Objectives 8
Official objectives from the NVIDIA outline, in outline order.
1.1 Describe the NVIDIA software stack used in an AI environment.CUDA, CUDA-X libraries, frameworks, NGC and NVIDIA AI Enterprise.
- What it means
- This objective covers the layers of NVIDIA software between the GPU and the application: the GPU driver, the CUDA platform, CUDA-X libraries such as cuDNN and NCCL, AI frameworks like PyTorch and TensorFlow built on them, containers from the NGC catalog, and the NVIDIA AI Enterprise suite that packages and supports this stack for production.
- What to understand
- Think of the stack bottom up. The driver exposes the GPU, CUDA gives a programming model, cuDNN accelerates neural network primitives, NCCL handles multi-GPU communication, and frameworks call these libraries so data scientists rarely write CUDA directly. NGC supplies tested containers and models. NVIDIA AI Enterprise adds enterprise support, security patching and certified deployment on mainstream servers and clouds.
- How it shows up
- A question names a library or layer and asks what it does, such as which component handles collective communication across GPUs.
- Common confusion
- Treating CUDA as an AI framework; it is the parallel computing platform that frameworks are built on.
1.2 Compare and contrast training and inference architecture requirements and considerations.Compute, memory, precision, latency and scaling differences between training and inference.
- What it means
- This covers why training and inference stress infrastructure differently. Training repeatedly passes large datasets through a model and updates weights, so it needs massive compute, large GPU memory and fast GPU-to-GPU links. Inference runs a trained model on new inputs and is judged on latency, throughput and cost per query.
- What to understand
- Training is compute- and memory-bound, runs for hours to weeks, uses mixed precision such as BF16 or FP8, and often spans many GPUs joined by NVLink and InfiniBand. Inference can run on fewer or smaller GPUs, benefits from lower precision such as INT8 or FP4, and is scaled out by adding replicas behind a load balancer. Large language model inference can still need multiple GPUs because of model size.
- How it shows up
- A workload description is given, and you identify whether it is training or inference and which resource matters most.
- Common confusion
- Assuming inference always needs the same multi-node cluster used for training.
1.3 Differentiate the concepts of AI, machine learning, and deep learning.The nested relationship between AI, ML and DL.
- What it means
- This objective covers the hierarchy of terms. Artificial intelligence is the broad goal of machines performing tasks that need human-like intelligence. Machine learning is the subset where systems learn patterns from data instead of following hand-written rules. Deep learning is the subset of machine learning that uses multi-layer neural networks.
- What to understand
- Every deep learning system is machine learning and every machine learning system is AI, but not the reverse. Classic machine learning, such as decision trees or linear regression, often relies on engineered features and runs well on CPUs. Deep learning learns features itself from raw data like images, audio and text, needs large datasets, and is where GPU acceleration delivers the biggest gains. Generative AI and large language models are deep learning.
- How it shows up
- A technique or system is described, and you choose the narrowest correct category.
- Common confusion
- Calling any data-driven model deep learning; only neural networks with many layers qualify.
1.4 Explain the factors contributing to recent rapid improvements and adoption of AI.Accelerated compute, data growth, algorithms and accessible tooling.
- What it means
- This covers why AI moved from research into mainstream use. NVIDIA frames it as several forces combining: GPU-accelerated computing that made training large networks practical, abundant digital data, algorithmic breakthroughs such as the transformer architecture, and software frameworks, pretrained models and cloud access that lowered the barrier to entry.
- What to understand
- No single factor explains the shift. GPUs cut training time from months to days, which let researchers iterate. Large datasets gave deep networks something to learn from. Transformers enabled large language models and generative AI. Open frameworks, model hubs and cloud GPU capacity let organizations adopt AI without building everything themselves. Be ready to identify which factor a scenario emphasizes.
- How it shows up
- A question asks which development most enabled a change, such as training far larger models in practical time.
- Common confusion
- Crediting only algorithms; without accelerated compute and large datasets, modern deep learning would not be practical.
1.5 Explain the key AI use cases and industries.Where AI is applied: healthcare, finance, manufacturing, retail, automotive and more.
- What it means
- This objective covers common AI applications by industry: medical imaging and drug discovery in healthcare, fraud detection and risk modeling in finance, predictive maintenance and visual inspection in manufacturing, recommendation and demand forecasting in retail, autonomous driving in automotive, and generative AI assistants across nearly every sector.
- What to understand
- Link each use case to the AI technique behind it. Image analysis uses computer vision, chatbots and document summarization use large language models, recommenders use deep learning on behavior data, and fraud detection uses anomaly detection on transactions. Knowing the technique helps you infer infrastructure needs, for example that video analytics at the edge favors efficient inference hardware.
- How it shows up
- A business problem is described, and you identify the AI use case or technique that addresses it.
- Common confusion
- Treating generative AI as the only AI use case; predictive and perception workloads remain widespread.
1.6 Explain the purpose and use case of various NVIDIA solutions.DGX, HGX, DGX Cloud, NIM, NeMo, Omniverse and industry platforms.
- What it means
- This covers what major NVIDIA offerings are for. DGX systems are integrated AI servers, and DGX SuperPOD is a reference architecture for large clusters. HGX is the GPU baseboard partners build servers around. DGX Cloud delivers NVIDIA infrastructure as a service. NIM microservices package optimized models for deployment, NeMo builds and customizes generative models, and Omniverse targets 3D simulation and digital twins.
- What to understand
- Associate questions test purpose, not configuration. Know the industry platforms by name: Clara for healthcare, DRIVE for autonomous vehicles, Isaac for robotics, Metropolis for vision AI, Riva for speech AI, and Morpheus for cybersecurity. Distinguish turnkey systems (DGX) from partner platforms (HGX) and from software (AI Enterprise, NIM, NeMo).
- How it shows up
- A need is described, such as deploying a speech assistant or building a factory digital twin, and you pick the NVIDIA solution.
- Common confusion
- Confusing DGX, an NVIDIA-built system, with HGX, the GPU board used in partner servers.
1.7 Describe the software components related to the life cycle of AI development and deployment.Data preparation, training, optimization, deployment and monitoring tools.
- What it means
- This objective follows a model from raw data to production. Data preparation can be accelerated with RAPIDS libraries. Training uses frameworks such as PyTorch, often from NGC containers, or NeMo for generative models. Optimization uses TensorRT. Deployment uses Triton Inference Server or NIM microservices, typically on Kubernetes, and monitoring watches latency, accuracy and drift.
- What to understand
- Map each lifecycle stage to its tooling and its goal. Containers make environments reproducible from a developer workstation to a cluster. Triton serves models from many frameworks with dynamic batching and concurrent model execution; NVIDIA now groups it under its Dynamo inference platform. NIM wraps an optimized model, runtime and standard API into one container. MLOps practices tie the stages together with versioning and retraining.
- How it shows up
- A lifecycle stage is described, and you choose the component used there, such as RAPIDS for data preparation.
- Common confusion
- Treating TensorRT as a model server; it optimizes a model, while Triton or NIM serves it.
1.8 Compare and contrast GPU and CPU architectures.Many parallel cores versus few powerful cores, throughput versus latency.
- What it means
- This covers why GPUs accelerate AI. A CPU has a small number of powerful cores with large caches, optimized for low-latency serial tasks and branching logic. A GPU has thousands of simpler cores that run the same operation across large blocks of data at once, which suits the matrix math at the heart of neural networks.
- What to understand
- CPUs remain essential for the operating system, data loading, orchestration and serial code, so AI servers pair both. GPUs add Tensor Cores, specialized units for mixed-precision matrix multiply, and high-bandwidth memory (HBM) to feed them. The trade-off is throughput versus latency: GPUs finish large parallel batches far faster, while CPUs respond faster to single, branch-heavy tasks.
- How it shows up
- A workload is described, and you decide whether its parallelism favors a GPU or its serial logic favors a CPU.
- Common confusion
- Thinking GPUs replace CPUs entirely; AI servers still rely on CPUs for control and data handling.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
AI, ML and DL hierarchy
Deep learning sits inside machine learning, which sits inside AI.
- Remember
- Choose the narrowest correct label.
- Tested as
- Definition questions.
- Trap
- Labelling every model as deep learning.
-
CUDA
NVIDIA's parallel computing platform and programming model for GPUs.
- Remember
- Frameworks build on CUDA; it is not itself a framework.
- Tested as
- Software stack questions.
- Trap
- Calling CUDA a deep learning framework.
-
cuDNN and NCCL
cuDNN speeds neural network layers; NCCL handles collective communication between GPUs.
- Remember
- NCCL is the multi-GPU, multi-node communication library.
- Tested as
- Stack component questions.
- Trap
- Swapping the two libraries' roles.
-
TensorRT
An SDK that optimizes trained models for fast inference.
- Remember
- It uses layer fusion and reduced precision.
- Tested as
- Inference optimization scenarios.
- Trap
- Using it to train models.
-
Triton Inference Server
Open-source server that hosts models from many frameworks with dynamic batching.
- Remember
- It serves; TensorRT optimizes.
- Tested as
- Deployment questions.
- Trap
- Confusing serving with optimization.
-
NVIDIA NIM
Prebuilt microservices that bundle an optimized model, runtime and industry-standard API.
- Remember
- Fastest path to deploy a supported model.
- Tested as
- Generative AI deployment questions.
- Trap
- Assuming NIM is a training tool.
-
NVIDIA AI Enterprise
Licensed, supported software suite for production AI.
- Remember
- It adds enterprise support and security updates.
- Tested as
- Stack and licensing questions.
- Trap
- Treating it as hardware.
-
NGC catalog
NVIDIA's hub of GPU-optimized containers, models and resources.
- Remember
- Source of tested framework containers.
- Tested as
- Lifecycle questions.
- Trap
- Confusing it with a cloud provider.
-
Tensor Cores
Specialized GPU units for mixed-precision matrix math.
- Remember
- They drive most deep learning speedups.
- Tested as
- GPU architecture questions.
- Trap
- Equating them with ordinary CUDA cores.
-
Training vs inference
Training builds the model; inference applies it to new data.
- Remember
- Training stresses compute and interconnect; inference stresses latency and cost.
- Tested as
- Architecture comparison questions.
- Trap
- Sizing inference like training.
-
Reduced precision
Formats such as FP8, INT8 and FP4 trade a little accuracy for speed and memory savings.
- Remember
- Lower precision is most common in inference.
- Tested as
- Optimization questions.
- Trap
- Assuming full FP32 is always required.
-
DGX vs HGX
DGX is an NVIDIA-built system; HGX is a GPU board partners build servers on.
- Remember
- Both use the same class of GPUs.
- Tested as
- Solution-matching questions.
- Trap
- Treating them as identical products.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
AI vs machine learning vs deep learning
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Artificial intelligence | Machines performing tasks that need human-like intelligence | Broadest category, includes rule-based systems | Assuming all AI learns from data |
| Machine learning | Systems that learn patterns from data | Often uses engineered features | Assuming it always needs GPUs |
| Deep learning | Multi-layer neural networks that learn features | Needs large data and heavy compute | Using it for any statistical model |
GPU vs CPU
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| CPU | Few powerful cores with large caches | Optimized for latency and branching | Thinking it is obsolete in AI servers |
| GPU | Thousands of cores running the same operation | Optimized for parallel throughput | Expecting it to speed serial code |
| Tensor Core | GPU unit for mixed-precision matrix math | Accelerates neural network layers | Treating it as a general CPU core |
NVIDIA lifecycle software
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| RAPIDS | GPU-accelerated data science libraries | Data preparation and analytics | Using it for model serving |
| TensorRT | Inference optimizer and runtime | Makes a trained model faster | Treating it as a server |
| Triton | Multi-framework inference server | Hosts and batches requests | Confusing it with TensorRT |
| NIM | Packaged model microservice with a standard API | Ready-to-run deployment unit | Assuming it trains models |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
CUDA is a framework
- What candidates pick
- Pick CUDA when asked for a deep learning framework.
- Why it looks right
- CUDA appears in every NVIDIA AI discussion.
- Why it's wrong
- CUDA is the parallel computing platform; PyTorch and TensorFlow are frameworks built on top of it.
Clue to watch for The question asks what data scientists use to define and train models.
-
Optimize equals serve
- What candidates pick
- Choose TensorRT to host models behind an API.
- Why it looks right
- TensorRT is closely tied to inference.
- Why it's wrong
- TensorRT compiles and optimizes a trained model; Triton or NIM host it and handle requests.
Clue to watch for The scenario mentions serving many requests or multiple frameworks.
-
Inference needs a training cluster
- What candidates pick
- Recommend the largest multi-node cluster for a modest inference service.
- Why it looks right
- Bigger seems safer.
- Why it's wrong
- Most inference is sized for latency and cost per query and scales out with replicas.
Clue to watch for The workload applies a finished model to new inputs.
-
All ML is deep learning
- What candidates pick
- Call a decision tree model deep learning.
- Why it looks right
- Deep learning dominates headlines.
- Why it's wrong
- Deep learning specifically means multi-layer neural networks; decision trees are classic machine learning.
Clue to watch for The described model uses engineered features or no neural network.
-
Algorithms alone drove adoption
- What candidates pick
- Credit only new algorithms for the AI boom.
- Why it looks right
- Transformers are well known.
- Why it's wrong
- Adoption came from GPU computing, large datasets, algorithms and accessible tools together.
Clue to watch for The question asks what made training large models practical.
-
DGX is a board
- What candidates pick
- Describe DGX as the GPU baseboard partners use.
- Why it looks right
- The names are similar.
- Why it's wrong
- HGX is the partner baseboard; DGX is NVIDIA's own complete system.
Clue to watch for The scenario mentions a server built by a third-party manufacturer.
Answer framework
Use these steps to reason through any Domain 1 concept or product question on NCA-AIIO.
- Decide whether the question is about a concept definition, a workload type, an architecture comparison or an NVIDIA product, since each has its own vocabulary.
- For definitions, place the technique in the AI, machine learning and deep learning hierarchy and pick the narrowest label that is still correct.
- For workloads, identify training versus inference and then ask which resource dominates: compute and interconnect for training, latency and cost for inference.
- For products, find the lifecycle stage first (data preparation, training, optimization, serving, packaging) and then pick the NVIDIA tool built for that stage.
- For architecture, check whether the work is massively parallel matrix math, which favors GPUs and Tensor Cores, or serial control logic, which favors CPUs.
Practice question
Want more like this? The NCA-AIIO practice test has questions for every domain.
Checklist
0 of 13
Tick each one you can do without notes. Your ticks stay in this browser.
Study strategy
Start with the definitions, because they are quick wins. Draw the AI, machine learning and deep learning circles and add an example to each, then write two or three lines contrasting GPUs and CPUs. These ideas carry through the whole exam.
Next, build a one-page map of NVIDIA software arranged by lifecycle stage: RAPIDS for data preparation, frameworks from NGC for training, NeMo for generative models, TensorRT for optimization, Triton and NIM for serving, and NVIDIA AI Enterprise as the supported wrapper. Add the CUDA, cuDNN and NCCL layers underneath. Most product questions are answered by finding the stage first.
Then make flashcards for NVIDIA platforms and their industries, since these are pure recall. Finish with training versus inference, because it links directly to the hardware sizing in AI Infrastructure. NVIDIA's free self-paced course, AI Infrastructure and Operations Fundamentals, covers this material in about seven hours. Pair it with short timed question sets, and review every miss by noting which lifecycle stage or definition you misread.
FAQ
How much of NCA-AIIO is Essential AI Knowledge?
Essential AI Knowledge is 38% of the exam, about 19 of the 50 questions, making it the second-largest of the three domains.
Do I need to write code for this domain?
No. The exam is associate level and tests recognition of concepts, products and architectures, not programming or command syntax.
What is the difference between TensorRT and Triton?
TensorRT optimizes a trained model for fast inference on NVIDIA GPUs. Triton Inference Server hosts models, including TensorRT engines, and handles requests and batching.
Why do GPUs accelerate deep learning?
Neural networks are dominated by matrix multiplication, which splits into many identical operations. GPUs run thousands of these in parallel and add Tensor Cores for mixed-precision math.
What are NVIDIA NIM microservices?
Containers that package an optimized model, an inference runtime and an industry-standard API so a team can deploy a supported model quickly on NVIDIA GPUs.
How deep should my knowledge of NVIDIA industry platforms be?
Know each platform's purpose and target industry, such as Clara for healthcare or Isaac for robotics. Configuration detail is beyond this exam.
Is generative AI covered in this domain?
Yes, as a use case and as context for products like NeMo and NIM. Expect questions on where it fits in the deep learning family rather than on prompt design.
Other Artificial Intelligence & Machine Learning certifications
Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.
Test your Domain 1 knowledge
You have covered AI fundamentals, GPU and CPU design, training and inference, and the NVIDIA software stack. Check whether you can match the right concept or product to a scenario under time pressure.
Keep going
- Training Training Built on your Weak areas
- Practice test Questions by domain, with explanations
- Mock exam Full length, timed like the real exam
- Readiness test Find your weakest domain in minutes
- Cheat sheet Key terms and formulas on one page
- Ask the NCA-AIIO AI tutor Get any concept explained your way