NCA-AMAD (NVIDIA-Certified Associate: AI Models and Application Development) is NVIDIA's associate exam for developers who build applications on top of AI models. NVIDIA describes its scope as "the full lifecycle of an AI application: selecting and customizing models, building retrieval and data pipelines, optimizing inference, deploying with NVIDIA AI software, and evaluating and improving systems in production." Agents are part of it too: how they plan, use tools and coordinate, along with the sizing and configuration decisions a deployment needs.
Exam Quick Facts
The NCA-AMAD article series
- 20 NCA-AMAD practice questions with explanations, spread across all eight sections
- The 4-week NCA-AMAD study plan, section by section
- NCA-AMAD practice tests: what is available and how to use it
Who NCA-AMAD is for
NVIDIA wrote the exam for "entry-level AI developers with a foundational understanding of how AI models are customized, served and consumed as applications." The certification page describes what those developers do: they select and evaluate pretrained models for a use case, customize them through fine-tuning or prompt engineering, deploy them with inference serving frameworks, and integrate them into enterprise applications.
That fits a software engineer who has built a chatbot or RAG assistant on a hosted API and now has to choose models, run them on GPUs and keep them working, and an early-career ML engineer who wants the application and serving side of the job on paper.
The exam page lists no formal prerequisites. The course NVIDIA recommends for the exam asks for basic machine learning and deep learning concepts, a basic understanding of computer hardware, basic Python and some experience with the command line.
Try 15 NCA-AMAD practice questions free
They come from the NCA-AMAD practice tests, and each option shows its explanation once you answer.
- 6 timed practice exams
- Every answer explained
- A score for each domain
- $19.99 one-time
How the questions read
The exam has 50 to 60 multiple-choice questions in one hour and is scored pass/fail. NVIDIA's exam details also list "Hands-on lab: Yes" without describing the lab, as of October 2026.
NVIDIA publishes sample questions for each section, and they show the style. Most are short situations that ask which technique, setting or NVIDIA product fits:
- A support agent keeps calling a database tool and stalls before it issues a refund. The fix NVIDIA's sample gives is a limit on tool-call iterations per request.
- A summarization service receives a long document with every request. Most of the GPU work per request goes into processing those input tokens.
- A 70-billion-parameter model is served on four NVLink-connected GPUs, and the goal is the fastest generation for each user. The answer is tensor parallelism.
The wrong options in these samples are real techniques and real NVIDIA products, used in the wrong place. Knowing what each tool does is the starting point; the exam then asks which one fits a given constraint.
The eight exam sections
NVIDIA weights the eight sections as follows and publishes a list of potential exam topics for each.
| Section | Weight |
|---|---|
| Agentic AI | 17% |
| Data Processing and Data Flywheel | 17% |
| Inference Concepts | 15% |
| Model Customization and Fine-Tuning Concepts | 14% |
| Models and Model Types | 11% |
| NVIDIA AI Enterprise and Blueprints Overview | 10% |
| Sizing and Configuration | 10% |
| Physical AI | 6% |
Agentic AI (17%)
Agentic AI shares the top weight with data processing. It covers how agents differ from plain generative AI and from rule-based systems, when a single model call is enough and when a multi-step workflow is worth its cost, and how agents reach tools and external systems through APIs and the Model Context Protocol (MCP). Reasoning patterns (ReAct, reflection, planner-executor), short-term and long-term memory, and single-agent versus multi-agent designs all appear. So do safety controls: guardrails, output filters, prompt-injection defenses, hallucination reduction and human-in-the-loop approval. NVIDIA names its own tools here: NeMo, the NVIDIA Agent Toolkit and NemoClaw. Measurement matters as well, with task-completion rate, token usage, latency and cost listed as the measures of agent performance.
Data Processing and Data Flywheel (17%)
This section follows data from raw collection to a model that keeps improving. Topics include feature engineering and feature stores, lineage and provenance for audits, ETL with Spark, Databricks, Kafka and NVIDIA RAPIDS, and cleaning and labeling text. Privacy has its own line: handling personal information, retention, data sovereignty, bias and misuse. NVIDIA also lists the difference between structured, semi-structured and unstructured data, data-quality problems such as missing values, duplicates and stale records, and why an evaluation built from your own application's data often says more than a general benchmark. The flywheel part asks how production feedback flows back into training, and what to monitor in production: accuracy, groundedness, hallucination rate and task completion.
Inference Concepts (15%)
Inference questions are about serving models fast and at a sensible cost. Expect the two phases of LLM inference (prefill, which processes the prompt, and decode, which generates tokens) and the trade-off between latency and throughput. Optimizations named in the topic list include quantization, KV and prefix caching, speculative decoding, tensor parallelism, disaggregated serving and batching, along with TensorRT-LLM and NVIDIA Dynamo. On the infrastructure side: model servers, gateways and routing, autoscaling, request queuing and cold starts. Questions also use the measures teams report, such as time to first token and throughput, and the drivers of inference cost: model size, token counts, traffic, hardware and latency targets.
Model Customization and Fine-Tuning Concepts (14%)
The core of this section is choosing between prompting, retrieval and training. It covers prompt engineering (system prompts, zero-shot and few-shot prompting, chain-of-thought, structured output), the full RAG workflow (ingestion, chunking, embeddings, vector indexing, hybrid search and reranking), and fine-tuning methods from SFT and PEFT through LoRA, QLoRA, RLHF, DPO, KTO, GRPO, distillation and instruction tuning. It also covers the risks of fine-tuning, catastrophic forgetting and overfitting, and how a tuned model reaches production: serving adapters or merging them, and validating with offline experiments and A/B tests. The blueprint files RAG under customization.
Models and Model Types (11%)
Here the exam checks that you know which kind of model does which job. NVIDIA lists neural networks versus classical machine learning, computer vision tasks (classification, detection, segmentation, OCR), and why parallel hardware suits neural networks. The non-language model types are embedding, reranking, multimodal, omni, vision-language and world foundation models. For language models: small versus large, reasoning models, base versus instruction-tuned, open versus proprietary, dense versus mixture-of-experts, and foundation versus task-specific. The mechanics of LLMs also appear (tokens, context windows, autoregressive generation, embedding space), as do NVIDIA's model families such as Nemotron and the criteria for choosing a model for a workload.
NVIDIA AI Enterprise and Blueprints Overview (10%)
This section is about NVIDIA's software for taking AI applications to production. It covers what NVIDIA AI Enterprise is for and where it runs, NVIDIA NIM microservices (how they package a model, accelerate inference and select a configuration for the GPU, and how they differ from serving a model yourself), and NVIDIA Blueprints as reference architectures that you match to a business scenario and build on with NIM. Deployment basics are included too: containers, Kubernetes, infrastructure as code, service meshes, and running the same stack on premises, in the cloud and at the edge. The AI factory architecture, with its layers from infrastructure to applications, closes the list.
Sizing and Configuration (10%)
Sizing and Configuration covers the operations side of a deployed AI application. Topics include model lifecycle and release management, retraining triggers and versioning; monitoring for drift, latency, throughput and health, including detecting SLA breaches and the measures MTTD and MTTR; and compute sizing (GPU memory, replicas, horizontal versus vertical scaling, concurrent users, latency targets). NVIDIA also lists AI factory operations with runbooks and playbooks, the parameters that change performance and cost (batch size, maximum sequence length, precision), Run:ai compute management, and observability: dashboards, tracing of prompts, agents and retrieval, logging, alerting, telemetry, and drift and anomaly detection.
Physical AI (6%)
The smallest section covers AI that acts in the physical world. It includes how physical AI differs from digital AI, NVIDIA's three-computer pattern (one computer to train, one to simulate, one on the robot), robotics and autonomous machines, and inference at the edge. Simulation gets the most attention: digital twins in NVIDIA Omniverse and Isaac Sim, synthetic data, domain randomization, the gap between simulation and the real world, and the flywheel that closes it. The model families named are NVIDIA Cosmos and Isaac GR00T.
Three questions from the NCA-AMAD practice tests
A manufacturer wants a knowledge assistant over 100,000 PDFs full of tables, charts and diagrams, with GPU-accelerated ingestion, hybrid search and options for multi-turn chat. The team wants a production-ready starting point instead of assembling components. Which blueprint fits?
[object Object]
Correct. NVIDIA describes the AI Blueprint for RAG as a production-ready, modular reference architecture for high-accuracy, high-performance RAG over multimodal enterprise data, covering ingestion, retrieval, reasoning and generation, with NeMo Retriever extraction for tables and charts and options such as hybrid search and multi-turn query rewriting.
A. AI-Q relies on Nemotron RAG retrieval, so document ingestion is still needed.
C. VSS analyzes video streams, and turning documents into video would lose their text and table structure.
D. The flywheel improves models from production logs; it is not a document assistant.
A team must choose between two models for a software agent that triages GitHub issues and opens fixes. Model A leads on MMLU, GSM8K and HumanEval. Model B scores lower there, yet resolves more issues in a trial run inside the team's real repository. Which evidence should weigh most?
[object Object]
Correct. Agent evaluation uses dynamic, end-to-end benchmarks such as SWE-bench, GAIA and WebArena, plus task success on your own scenarios. Model B's results in the real repository are the closer evidence.
A. Parameter count and window length describe the model design. Only measured task results show which model completes the team's tasks.
B. Re-weighting static scores still measures isolated capabilities. It does not test planning, tool calls or recovery inside a real environment.
D. NVIDIA notes that a high MMLU score is a prerequisite for a reliable agent but does not guarantee one. Static benchmarks test the model in isolation.
Every request to a company assistant starts with the same 3,000-token system prompt and tool definitions, followed by a short user question of about 100 tokens. Time to first token is mostly spent processing that identical prefix on every call. Which optimization targets this directly?
[object Object]
Correct. TensorRT-LLM stores the KV cache blocks of completed requests and reuses them for later requests that share the same prefix. Only the new question tokens need prefill, which cuts TTFT and GPU work.
A. Shortening the prompt removes instructions the assistant relies on, and tool definitions are needed as input.
B. Lower precision makes each pass cheaper, yet the identical prefix is still recomputed on each call.
C. Batching processes the same prefix many times in parallel. The repeated work remains.
Full NCA-AMAD set: 6 timed exams, every answer explained, $19.99 one-time.
See the practice testsNCA-AMAD or NCA-GENL?
Both are NVIDIA associate exams for developers, at the same price and length. Their blueprints point in different directions.
| NCA-GENL | NCA-AMAD | |
|---|---|---|
| Sections | Core Machine Learning and AI Knowledge, Software Development, Experimentation, Data Analysis and Visualization, Trustworthy AI | Agentic AI, Data Processing and Data Flywheel, Inference, Customization, Models, NVIDIA AI Enterprise and Blueprints, Sizing and Configuration, Physical AI |
| Center of gravity | Machine learning and LLM fundamentals, software development and experimentation | Building, serving and running AI applications and agents on NVIDIA's stack |
| Sections on serving, sizing and NVIDIA deployment software | None by name | Inference (15%), Sizing and Configuration (10%), NVIDIA AI Enterprise and Blueprints (10%) |
If you are learning how language models work and building your first experiments, NCA-GENL matches that stage. If your work already involves putting models behind applications (choosing a serving setup, adding retrieval, sizing GPUs, running agents), NCA-AMAD tests more of what you do. Either one leads naturally to the professional exams for agentic AI (NCP-AAI) and LLMs (NCP-GENL). The NCA-GENL complete guide covers the other exam.
How to prepare
Start from NVIDIA's topic list. The certification page lists potential exam topics and suggested reading for every section, mostly NVIDIA technical blog posts and documentation. Those readings define the vocabulary the questions use, for example the NeMo Guardrails rail types, LLM inference benchmarking concepts and the data flywheel.
Take NVIDIA's recommended course if you can. AI Model and Application Development Fundamentals is an eight-hour, beginner-level, instructor-led workshop (priced on request as of October 2026). It walks the lifecycle with hands-on work: preparing large datasets with CUDA-X libraries, building models with NVIDIA NeMo, deploying and optimizing inference with NVIDIA Dynamo, and running workloads with Kubernetes, Slurm and an observability stack.
Build the systems the topic list names. Preporato Pro includes hands-on labs for those skills: agents with tools, MCP, memory and NeMo Guardrails; RAG with chunking, hybrid search and reranking; LoRA, QLoRA and DPO fine-tuning; LLM serving with vLLM, batching, quantization and load tests; data curation and PII redaction; and tracing and GPU monitoring. They run in the browser on hosted NVIDIA NIM endpoints, NVIDIA GPU pods and Kubernetes, with nothing to install and an automatic check at every step. Physical AI has no lab yet. NVIDIA does not describe the exam's hands-on lab, so use the labs to practice the skills in its topic list.
NCA-AMAD skills you can build in the browser
One lab each for agents, RAG, fine-tuning, serving, data and tracing. The first two are free with an account; the rest come with Pro.
Build an Agent From Scratch: The Tool Loop Without a Framework
Hand-roll a tool-using agent on the chat completions API and nothing else: parse the model's tool calls, dispatch them to plain Python functions, close the loop with correctly threaded tool results, add stop conditions and budgets, turn every failure into a tool result the model can read, put an approval guardrail in front of a side-effecting refund tool, control context with truncation and a trace, and finish with an evaluation run that scores the agent on a fixed question set.
Build a RAG Pipeline with NVIDIA NIM
Build a complete Retrieval Augmented Generation pipeline — from document chunking to vector search to an agent that answers questions from your knowledge base.
Fine-Tune an LLM with LoRA and QLoRA (Jupyter)
Fine-tune Meta Llama 3 8B on a custom instruction dataset using LoRA and QLoRA. Learn parameter-efficient fine-tuning from data preparation through evaluation — the #1 most demanded AI skill.
vLLM Production Serving: PagedAttention, Continuous Batching, Prefix Caching
Stand up vLLM and measure the three features that make it the de-facto inference server: PagedAttention's KV-cache capacity, continuous batching throughput, and prefix caching speedups. Then write the production spec — server args, Kubernetes deployment, monitoring, autoscaling.
Curate a Fine-Tuning Dataset: Deduplicate, Filter, LLM-Judge and Balance Instruction Data
Turn a raw 260-row helpdesk export into a clean instruction-tuning set. Profile it, strip empty, truncated, boilerplate and PII rows with rules, remove near-duplicates with shingles and Jaccard similarity, grade what rules cannot see with an LLM judge calibrated against human labels, then cap categories and write train/eval splits with a data card.
LLM Observability: Trace, Cost and Debug a RAG Assistant
Instrument a real RAG assistant with traces from scratch: nested spans around retrieval and generation, token usage and cost attributed to every call, a latency and cost report, a triage tool that finds retrieval misses from trace evidence, PII redaction before anything is written, sampling that keeps every error and slow request, a cost budget alarm, and a CI gate that blocks a prompt change that makes the assistant slower or dearer.
Answer NVIDIA's sample questions. The certification page has sample questions for each section in a Learning Mode that shows the correct answer, an explanation and a link to read more. They are the best guide to the style of the real exam.
Practice under time. Sixty minutes for up to 60 questions leaves about a minute per question. Timed practice tests show which sections cost you points while there is time to fix them. The free NCA-AMAD practice questions on the certificate page come from the full preporato.com practice tests, and the 4-week study plan puts the steps above into a schedule.
Frequently asked questions
How hard is the NCA-AMAD exam?
It is an associate exam, and its difficulty comes from breadth. Eight sections run from agents and data pipelines through inference, fine-tuning, model types and NVIDIA's software stack to deployment sizing and physical AI. The questions are short scenarios, so you need to know which technique fits a situation as well as what each term means.
Are there prerequisites for NCA-AMAD?
The certification page lists none. NVIDIA describes the target candidate as an entry-level AI developer, and its recommended course asks for basic machine learning and deep learning concepts, basic hardware knowledge, basic Python and some command-line experience.
What is the passing score for NCA-AMAD?
NVIDIA scores the exam pass/fail and does not publish a passing score, as of October 2026. Aim for consistent results of 80% or more on timed practice tests before you book, so that a hard draw of questions on the day still leaves a margin.
Does NCA-AMAD include a hands-on lab?
NVIDIA's exam details list "Hands-on lab: Yes" and, as of October 2026, do not describe the lab's format. The recommended course is hands-on, so it is the most direct preparation for practical work NVIDIA may ask for. On preporato.com, Pro includes hands-on labs for the skills in the topic list, from agents and RAG to fine-tuning, serving and monitoring.
How long does it take to prepare for NCA-AMAD?
For someone who already builds with language models, our study plan assumes four weeks of part-time study. Plan for longer if serving models, GPUs or agents are new to you.
How long is the NCA-AMAD certification valid?
Two years from issuance. You recertify by taking the exam again. Passing earns a digital badge and an optional certificate.
Where can I try NCA-AMAD questions for free?
NVIDIA's sample questions are on the certification page. The certificate page on preporato.com has free NCA-AMAD practice questions with explanations, and the NCA-AMAD practice questions article walks through 20 more.
Sources: