Domain 1: Experimentation
- An experiment differs from simply trying something by having a stated expectation the result can support or contradict. If no outcome would have changed your mind, nothing has been learnt whatever the number says.
- A baseline is what a new method has to beat before it is interesting. Without one, an impressive-looking figure has nothing to be impressive relative to.
- The evaluation portion has to be kept away from every decision made while developing. Choosing a threshold, a stopping point or a prompt while watching that portion turns it into training data and inflates the figure it reports.
- Data with a time order splits by time, not at random. A random split lets the model see the future of a sequence whose past it is being asked to predict.
- Automatic metrics for generated output measure agreement with a reference, which penalises a correct answer expressed in different words and rewards a wrong answer expressed in familiar ones.
- Human evaluation needs written criteria, more than one rater and a measure of how often the raters agree. Ratings without an agreement figure cannot be distinguished from noise.
- A single run gives a single draw. Reporting the best of several runs, or comparing two methods measured once each, describes the luck rather than the methods.
- An ablation removes one component and re-measures. It is the only way to attribute a gain to the part of a change you think caused it.
Domain 2: Core Machine Learning and AI Knowledge
- Supervised learning fits to labelled examples, unsupervised finds structure without labels, and reinforcement learning improves through feedback on actions taken. Most generative pretraining is self-supervised - the data supplies its own target.
- The loss measures how wrong the model currently is, and fitting moves each parameter a little in the direction that reduces it. Everything about learning rates, batching and convergence follows from that.
- Under-fitting shows as poor performance on training and held-out data alike; over-fitting shows as excellent training performance with poor held-out performance. The two call for opposite responses.
- Attention lets every position relate to every other position, which is what gives a transformer its context. The cost of that relating climbs steeply as the context grows.
- Diffusion generates by starting from noise and removing a little of it at each step. Fewer steps finish sooner and leave more imperfection behind.
- Performing diffusion in a compressed space rather than on pixels cuts the computation enormously, at the cost of needing something to compress into that space and something to expand back out.
- An embedding places an item at a position whose proximity to other positions means similarity. A shared space for two modalities is what makes a sentence able to retrieve a picture.
- Adapting a model by fine-tuning changes its weights and suits a behaviour; supplying retrieved context changes nothing and suits a fact. A fact that changes weekly belongs in the prompt, not in the weights.
Domain 3: Multimodal Data
- The pairing is the dataset. A picture and a caption that do not correspond teach the model a relationship that is not there, and no amount of volume repairs a systematic pairing error.
- Captions harvested from the web are plentiful and loosely related to their images; captions written deliberately are accurate and scarce. Most systems use a great deal of the first and a little of the second.
- Duplicates that straddle a split leak the evaluation into the training. Near-duplicates leak just as effectively and are far harder to find.
- Every modality needs preparation, and the same preparation has to run at inference as at training. A step applied in one place and not the other produces a model that tests well and performs badly.
- Images are resized to a fixed shape, which means either distorting the proportions or cropping away part of the picture. Which is preferable depends on whether the subject reaches the edges.
- Audio is usually presented as frequency over time. A short analysis window locates events precisely in time and blurs their frequency; a long one does the reverse.
- Normalisation statistics are computed on the training portion and applied to every portion. Computing them across everything leaks the evaluation into the preparation.
- A model given several modalities will use the most informative one and may ignore the rest. Removing a modality and watching the performance not move is how you find out.
Domain 4: Software Development
- A call to a hosted model is a network call, with every network failure mode attached. Timeouts, retries with growing delays, and a limit on attempts are not optional refinements.
- A timeout says the caller stopped waiting, not that the work failed. A retry without an idempotency key can pay for the same work twice.
- User text pasted into a prompt template can carry instructions. Marking the untrusted portion helps; limiting what the application will act on from the output helps more.
- The prompt, the model version and the surrounding code together determine behaviour. All three need versioning, review and a way back.
- Test the surrounding code against a fixed stand-in for the model, and test quality against the real model on a schedule. A test that passes sometimes trains everybody to ignore failures.
- Batching raises throughput and makes the first request in a batch wait. The largest batch size and the longest wait are the two settings that govern the trade.
- Ask for structured output and still validate it. On a validation failure, retry once with the failure described, or fall back to something the application knows how to handle.
- Pin dependencies, keep credentials outside the repository, and ship the same artefact to test and production with configuration supplied from the environment.
Domain 5: Data Analysis and Visualization
- Read real rows before computing anything. Encoding problems, wrong units and empty fields surface faster in ten examples than in any summary.
- A mean says nothing about spread, and a mean sitting between two clusters describes a value almost nobody has. A histogram shows what no single number can.
- Why a value is missing matters as much as how often. A value missing because of what it would have been makes every surviving row a biased selection.
- Two fields moving together is consistent with either causing the other or with something causing both. Only intervention or a clear ordering in time strengthens the claim.
- A relationship that holds in every subgroup and reverses overall is driven by the mix. Report the subgroups and state their composition.
- Bars are read as lengths and need a zero baseline; a line showing movement may reasonably omit zero. Two vertical axes let almost any pair be made to look connected.
- A rate from a small group swings wildly on a single observation, which is why the highest and lowest rates in any ranking are usually the smallest groups.
- Stopping a comparison as soon as it looks favourable reports the wandering rather than the effect. Decide the duration before the first figure arrives.
Domain 6: Performance Optimization
- Measure before changing anything. A portion taking a tenth of the time cannot yield more than a tenth of an improvement however elegantly it is rewritten.
- An accelerator sitting idle is waiting on something other than the computation, and the usual answer is that data is not being prepared and delivered quickly enough.
- Work is submitted asynchronously, so a timer stopped at submission measures the submission. Discard the first iterations, then repeat and report the spread.
- Halving the bits halves the memory and usually speeds the arithmetic, at the cost of precision. For training, range matters more than fraction bits, which is why loss scaling exists.
- Quantising a trained model to eight bits gives roughly a quarter of the memory and an accuracy cost that has to be measured on the real task rather than assumed.
- During generation, memory grows as the response lengthens because earlier positions are retained. Length and concurrency together decide how many conversations a device can hold.
- Single-token generation reads every parameter to do very little arithmetic, which makes it memory-bound. Serving several requests at once amortises that read across them.
- Distillation, pruning, a smaller model for a narrow task, and routing easy requests away from the large model all trade a little quality for a great deal of cost.
Domain 7: Trustworthy AI
- A model performs worst for the groups its data barely represented, and a single overall figure conceals that. Report performance separately for each group of interest.
- Different definitions of fairness can be mathematically incompatible. The one you chose, and why, belongs in writing where somebody can disagree with it.
- Content repeated many times in training data can be reproduced in an output long afterwards. De-duplication and identifier removal act before any harm is possible.
- A prompt sent to a hosted model leaves the sender's control, and what may be placed in one has to be decided deliberately rather than left to each user.
- Publicly accessible is not the same as permitted. A record of where each portion of training data came from, and under what terms, is what answers the question later.
- Asked about something absent from its training, a model produces a fluent answer that may be entirely invented, and a confidence it states is generated in the same way the answer was.
- A citation that does not support the sentence it follows is worse than none, because it borrows credibility it has not earned. Check the passage, not the formatting.
- Responsibility rests with whoever chose to deploy the system. A human review that never disagrees with the model has stopped being a control.
NVIDIA-Certified Associate exam tips
- Experimentation is the largest domain at 25 percent, and most of its questions are about why a reported number should not be believed rather than about how to compute one. If you are short of time, study this domain first.
- Many stems describe a situation and ask what follows from it. Read for what has actually been observed rather than for what might also be true, because several options are usually real concepts that do not answer the question asked.
- Where a question offers two options that differ by one word - rises or falls, more or fewer, before or after - the direction is the thing being tested. Work out which way the effect runs before looking at the wording.
- Multi-select questions state how many answers to choose. Choosing the right number matters as much as choosing the right options, and partial credit should not be assumed.
- Trustworthy AI is only 5 percent of the exam but the questions are unambiguous once you have the principle. It is the cheapest domain to secure, so do not leave it unread.
- Performance Optimization questions are usually about which resource is binding - memory or arithmetic, capacity or bandwidth. Identify the constraint and most options eliminate themselves.
- Software Development questions assume a hosted model behind a network call. Anything that would be true of any remote service - timeouts, retries, rate limits, versioning - is usually true here too.
- The exam is 60 minutes for up to 60 questions, so a minute each. Answer what you know quickly and leave the reasoning time for the scenario-style items rather than spending it on definitions.
Study guide FAQ
What is the NCA-GENM exam format?
50 to 60 multiple-choice and multiple-select questions in 60 minutes, delivered online and remotely proctored. NVIDIA does not publish a numeric passing score, so the result is reported as pass or fail. The certification is valid for two years.
How does NCA-GENM differ from NCA-GENL?
NCA-GENL covers generative AI with large language models. NCA-GENM covers generative AI across several modalities at once - images, audio, video and text together - and adds a Multimodal Data domain and a Performance Optimization domain that GENL does not have. The two overlap on core machine learning knowledge, experimentation and trustworthy AI.
Do I need to know NVIDIA-specific products to pass?
The published content breakdown is written in general terms - experimentation, multimodal data, software development, optimization - rather than around named products. Familiarity with the NVIDIA stack helps with context, but the questions test the underlying practice rather than product configuration.
How much mathematics does the exam require?
Very little computation. You need to reason about what a loss is, why a distribution matters, how memory scales with batch size or context length, and why a difference smaller than the run-to-run variation is not evidence. Nothing requires working an equation on paper.
Is programming experience required?
Software Development is 15 percent of the exam and assumes you have called a model from application code, but the questions are about the shape of that code - retries, validation, versioning, testing - rather than about syntax in any particular language.
How should I use this question bank?
Work through it by domain first, reading the explanation on every question including the ones you answered correctly, because the explanations state why each wrong option is wrong. Then take full weighted mocks, which draw questions in the same 25/20/15/15/10/10/5 proportion the real exam uses.