Six phases, ordered so each one builds on the last. Every phase tells you how the exam questions that material, gives you a practice exercise, and ends with self-checks you should be able to answer without notes. The exercises are optional if you are short on time; the self-checks are not.
To study for the NVIDIA NCA-GENL certification, work through the five exam domains in order of weight rather than the order they are listed. Start with Core Machine Learning and AI Knowledge (30% of the exam), which covers transformer architecture, embeddings, retrieval-augmented generation and prompt engineering, and underpins everything else. Move to Software Development (24%) for model serving with Triton and TensorRT, quantization, batching and monitoring. Then Experimentation (22%), which is almost entirely about evaluation metrics, A/B testing and reinforcement learning from human feedback. Finish with Data Analysis and Visualization (14%), including the RAPIDS and cuML GPU stack, and Trustworthy AI (10%), where bias detection, privacy and the NIST AI Risk Management Framework carry the most marks. Build a small evaluation set early and reuse it, since measurement appears in three of the five domains. Most candidates with working Python and some machine-learning background need 30 to 45 hours of focused study.
Study time should follow exam weight. The five domains and their share of the marks, heaviest first — this is the order the six phases below follow.
The algorithms, architectures and techniques behind machine learning and large language models — and the judgement to pick between them for a stated task.
Building, serving, optimising and monitoring the software around a model — the engineering that turns a checkpoint into a service.
Performing, evaluating and interpreting experiments — including model evaluation and the use of human subjects in labeling and reinforcement learning from human feedback.
Inspecting, cleaning and transforming data, and presenting what you find — with a strong emphasis on GPU-accelerated tooling.
Building AI systems that are ethical, fair, private, transparent and verifiable — and knowing which control evidences each of those claims.
Each phase names the domains it covers, how the exam questions that material, a practice exercise, and the self-checks that tell you whether to move on.
Be able to explain what a language model is trained to do, how a transformer processes a sequence, where a model's knowledge lives, and what shape the data is in — well enough that every later topic has somewhere to attach.
A written account, in your own words, of next-token prediction, the attention mechanism, and how parameter count relates to memory footprint — with one worked arithmetic example.
Build a retrieval-augmented pipeline and reason about every stage: chunking, embedding, indexing, retrieval, generation — and know which stage produced a bad answer.
A working RAG pipeline over a corpus you chose, plus a list of failures where each bad answer is attributed to a specific stage.
Evaluate models and changes like an experiment: a held-out set, one variable at a time, a stated metric, and a result you would defend in a review.
A frozen evaluation set of 20 or more items for your Phase 2 pipeline, scoring retrieval quality and answer quality separately, plus one A/B comparison you ran and interpreted.
Inspect and prepare data competently, compare models with the right statistic, and know where RAPIDS and cuML fit.
A short data-quality report on your Phase 2 corpus: length distribution, duplicates, topic balance, and one chart you would put in front of a senior engineer.
Deploy a model-backed service and change its cost or latency deliberately, with a measurement to show for it.
A served model with monitoring on it, plus one optimization — quantization, batching, or a change of serving stack — whose effect you measured against your Phase 3 evaluation set.
Attach a concrete control to every principle, so you can name the mechanism and the evidence rather than describing an ideal.
A one-page trust checklist for your own service: each principle, the control that implements it, and the artifact that evidences it.
The same six phases on a calendar, at three intensities. Pick the one that matches the time you actually have — total hours matter less than leaving the judgement-shaped material time to settle.
You work with LLMs already and need the credential soon. Assumes you can give it two to three hours on weekdays and a full day each weekend.
The default recommendation. Around ten hours a week leaves room for the material to settle between sessions, which matters more than total hours for the judgement-shaped questions.
You know Python but have not built an LLM system. The extra weeks go on Phases 1 and 2, where someone new to the material needs the time — not on the smaller domains.
Everything the plan covers, grouped by domain. Titles that link are published guides; the rest are on the way.
30% · 10 topics · 3 published
A large language model is trained to do exactly one thing — predict the next token given every token before it — and every capability it appears to have, from answering questions to writing code, is that single objective applied repeatedly. Because the correct next token already sits in the training text, the objective is self-supervised: it needs enormous amounts of text but no human labels.
An LLM's parameters are the numbers learned during training — the weight matrices and biases inside every layer — and they are the only place the model stores anything it "knows." Nothing in a prompt, a retrieved document or a system message ever becomes a parameter, which is why prompting and RAG change behaviour and output while only fine-tuning changes what the model knows.
Data moves through a transformer as a three-dimensional tensor shaped (batch, sequence, hidden) — how many texts at once, how many tokens each, and how many numbers represent each token. Every layer in the stack consumes that shape and returns it unchanged, which is why the triple is the single vocabulary you need to read a model config, a memory estimate, or a shape-mismatch error.
The densest subject on the exam, and the foundation for most of Domain 1.
The one piece of linear algebra that has to be airtight — six later topics reduce to it.
24% · 6 topics
22% · 7 topics
Re-run in every later phase — the instrument that makes the rest measurable.
14% · 5 topics
10% · 6 topics
How people waste weeks preparing for this exam, and what to do instead. These are process errors rather than gaps in the material.
Drill only after you have worked a phase. A practice test tells you where you are weak; it does not teach you the material, and answering the same bank repeatedly teaches the bank rather than the subject.
Follow weight instead. Core Machine Learning and AI Knowledge is 30% of the paper and everything else builds on it, so starting anywhere else means learning the later domains twice.
Give each a focused afternoon. Together they are almost a quarter of the exam and they are the cheapest marks available anywhere on it — an hour on RAPIDS and cuML, and an hour on the NIST AI RMF functions, both pay back immediately.
That advice holds for Software Development onwards and is wrong for Domain 1, where attention and transformer mechanics run through most of the questions. Learn what each mechanism does and why it exists.
Build a small pipeline and break it on purpose. The questions ask you to attribute a bad answer to chunking, retrieval or generation, and that diagnosis is very hard to fake from reading alone.
Rehearse them. A quarter of the paper needs exactly two selections with no partial credit, so the habit of counting the required answers before reading the options is worth practising until it is automatic.
Track per-domain results and study the lowest. A strong average routinely hides one domain you would fail standing alone — most often Experimentation.
Technique rather than content — how to spend the hour, how to handle the questions that give no partial credit, and how to know when you are ready.
The paper can run to 60 questions in 60 minutes, so one minute each is the pace to rehearse. Flag anything still unresolved at 90 seconds and move on — a question you are stuck on costs you two you would have answered.
Multi-answer questions are phrased "Select TWO" and always need exactly two selections — never one, never three. Scoring is all-or-nothing, so a single correct pick earns the same as a blank. Check the required count before you read the options.
Applied questions name a constraint: latency, memory, cost, no retraining, must cite sources. Two options are usually right for the subject and wrong for the constraint. Find the constraint clause first, then eliminate against it.
Anything phrased around out-of-date facts, needing citations, or removing one user's data is a retrieval question, not a fine-tuning question. This single distinction is worth a disproportionate number of marks.
Triton serves, TensorRT optimises, NeMo customises, NeMo Guardrails constrains, RAPIDS and cuML accelerate dataframes and classical ML. One clear sentence for each is enough to pick the right option.
A strong overall average can hide a domain you would fail on its own. Track results per domain and keep studying the lowest one — that is a better readiness signal than any single score.
The questions people ask most often when planning their preparation.
Study the five domains in order of weight, not the order they are published in. Begin with Core Machine Learning and AI Knowledge, because transformers, embeddings, retrieval and prompting reappear in every other domain. Then cover Software Development, Experimentation, Data Analysis and Trustworthy AI. Build a small evaluation set early and reuse it throughout, since measurement is tested in three of the five domains.
Most candidates need 30 to 45 hours of focused study. Expect the upper end if you have never built a retrieval pipeline or deployed a model, and the lower end if you work with large language models daily. Experimentation usually takes longer than people plan for, because it is tested more deeply than its name suggests.
Yes, if you already write Python and understand neural networks. Two weeks at two to three hours a day covers the ground. Prioritise transformer architecture, the retrieval pipeline and evaluation metrics — those three subjects account for the majority of the questions. Skip the optional practice exercises and use the self-checks and timed practice questions instead.
Transformer architecture and the attention mechanism. It is the single most heavily questioned subject on the exam and every later topic assumes it. Once you can explain what self-attention computes and why it replaced recurrence, move on to embeddings and retrieval-augmented generation.
Experimentation catches most candidates out. It is worth 22% of the exam and focuses almost entirely on evaluating models — choosing metrics, designing A/B tests, and understanding reinforcement learning from human feedback. Candidates who prepare it as general data analysis lose marks, because charting and data-mining skills are not what it asks about.
No, but you do need to know the product landscape. Triton serves models, TensorRT optimises them, NeMo customises them, NeMo Guardrails constrains their outputs, and RAPIDS and cuML accelerate dataframes and classical machine learning. One clear sentence about each is enough to answer the questions that name them.
Not on its own. Practice questions show you where the gaps are but rarely teach the reasoning behind an answer, and the exam favours judgement over recall. Use them after each study phase to find your weakest domain, then go back to the material for that domain rather than repeating the questions.
Two to three hours a day over three to four weeks suits most people, which fits the 30 to 45 hour range comfortably. Shorter daily sessions with a self-check at the end work better than long weekend blocks, because the material is broad and benefits from spaced review.
Core Machine Learning and AI Knowledge first, then Embeddings and retrieval, then Experimentation, then Data Analysis, and Trustworthy AI last. This follows exam weight and dependency: the fundamentals support everything, and the two lightest domains are quick to prepare once you understand the rest.
When you score consistently across all five domains with none lagging, and you can answer the self-checks in this guide without notes. NVIDIA does not publish a passing score, so per-domain consistency is a better signal than any overall percentage — a strong average can hide one domain you would fail on its own.
Practise by domain to find which phase you have not really finished, or sit a full-length timed mock at exam pace.