NVIDIA Certified Associate: Generative AI LLMs NCA-GENL Cheatsheet
Everything worth re-reading in the hour before you sit NCA-GENL, on one page. No explanations — just the facts, the weights, the traps and the judgement calls. Print it or keep it open beside your last practice run.
Exam facts
- Credential
- NVIDIA-Certified Associate: Generative AI LLMs
- Exam code
- NCA-GENL
- Level
- Associate (entry-level)
- Duration
- 60 minutes
- Questions
- 50–60
- Passing score
- Not published
- Price
- $125 USD
- Delivery
- Online, remotely proctored
- Validity
- 2 years — recertify by retaking
- Question format
- Multiple choice
Domain weights
- Core Machine Learning and AI Knowledge30%
- Software Development24%
- Experimentation22%
- Data Analysis and Visualization14%
- Trustworthy AI10%
Study time follows weight. The top two domains are more than half the exam.
Exam-day tactics
- Pace at 60 seconds a question. 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.
- Every multi-answer question wants exactly two. 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.
- Eliminate on the constraint, not the topic. 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.
- Freshness and attribution mean retrieval. Anything phrased around out-of-date facts, needing citations, or removing one user's data is a retrieval question, not a fine-tuning question.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.
- Know each NVIDIA product by its job. 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.
- Measure readiness by your weakest domain. 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.
One trap per domain
Advice that this exam is "all applied, no theory" is wrong here. Attention and transformer mechanics run through most of this domain, more heavily than any other subject anywhere on the paper. You will not be asked to derive the maths, but you do need to explain what a mechanism does, why it exists and how it fails. Learn the architecture properly first, then layer the selection decisions on top.
The phrase "write software components or scripts" makes this sound like a coding exam. It is not. Almost nothing here tests Python syntax; the questions concentrate on the serving and optimisation stack — Triton, TensorRT, vLLM, quantization, batching, monitoring. Prepare by learning what each tool is for and where it sits in a pipeline, not by drilling code.
This domain is worth more than one exam question in five, and it is where under-prepared candidates lose the most marks — partly because the name suggests broad data work when the questions are tightly focused on evaluating models. Build a real evaluation set, learn the metrics properly, and understand RLHF and A/B testing. Charting and data-mining skills earn nothing here.
This is the lightest domain, and generic pandas-and-matplotlib knowledge will carry most of the concepts — but not the tooling questions. NVIDIA is specific here: know what RAPIDS and cuML do and when you would reach for them. An hour on the GPU data-science stack is the highest-yield hour available anywhere on this exam.
The smallest domain is also the most underestimated. It sounds like a policy essay and is tested as engineering: expect to identify the specific control, framework function or evaluation that implements a principle. Learn the NIST AI RMF functions by name and know NeMo Guardrails as a product, and you will pick up marks most candidates leave behind.
Where the questions concentrate
The three subjects that carry the most marks in each domain.
Core Machine Learning and AI Knowledge30%
- Transformer architecture and the attention mechanism
- Embeddings
- Retrieval-augmented generation end to end
Software Development24%
- Inference serving and memory management
- Monitoring and observability of a live service
- NVIDIA Triton Inference Server
Experimentation22%
- Evaluation metrics, well ahead of anything else in this domain
- Reinforcement learning from human feedback
- A/B testing
Data Analysis and Visualization14%
- GPU-accelerated data science with RAPIDS and cuML
- Text preprocessing for model input
- Tokenization and text representation as a data-preparation concern, not just a modelling one
Trustworthy AI10%
- Bias, the largest subject in this domain by a wide margin
- Privacy and consent as engineering constraints
- The NIST AI Risk Management Framework and its four functions
Every NVIDIA tool by its one job
Most wrong options are a real product doing a job that belongs to a different one.
| Tool | Its job | Not for |
|---|---|---|
| Triton Inference Server | Serves models in production — multi-framework, multi-model, with batching and versioning | Optimising the model itself. Triton runs what you give it. |
| TensorRT / TensorRT-LLM | Compiles and optimises a model for a target GPU before serving | Serving. It produces an optimised engine; something else hosts it. |
| vLLM | High-throughput serving with paged attention for KV-cache memory | Training or fine-tuning of any kind. |
| NeMo | Builds and customises models — training, fine-tuning, alignment | Runtime policy enforcement. That is Guardrails. |
| NeMo Guardrails | Constrains a deployed system at runtime, with an auditable log | Removing bias from a model. It bounds behaviour; it does not retrain. |
| RAPIDS (cuDF) | GPU-accelerated dataframes, with a pandas-shaped API | Small datasets, where transfer overhead exceeds the speed-up. |
| cuML | GPU-accelerated classical machine learning, mirroring scikit-learn | Deep learning. That is PyTorch or TensorFlow. |
| Vector database | Stores embeddings and answers nearest-neighbour queries at scale | Generating embeddings. An embedding model produces them. |
Metrics, and what each one hides
The exam asks which metric answers which question. The blind spot is the half most candidates cannot state.
| Metric | Measures | Use when | Blind spot |
|---|---|---|---|
| Perplexity | How surprised a model is by text, derived from loss | Comparing two language models on the same corpus | Says nothing about accuracy, usefulness or safety |
| BLEU | N-gram overlap with a reference, precision-oriented | Translation, or any task with a tight reference answer | Punishes a correct paraphrase that uses different words |
| ROUGE | Overlap with a reference, recall-oriented | Summarization, where coverage matters more than precision | A summary can score well while being incoherent |
| Precision | Of what you flagged, how much was right | A false positive is the expensive error | Ignores everything you missed entirely |
| Recall | Of what was there, how much you found | A false negative is the expensive error | Flagging everything scores perfectly |
| F1 | Harmonic mean of precision and recall | Both errors matter, or the classes are imbalanced | One number hides which of the two is failing |
| Cross-entropy loss | Distance between predicted and actual token distributions | Monitoring training convergence | Falling loss can still mean overfitting |
| Explained variance | Share of variance in the target the model accounts for | Comparing regression models | Meaningless for classification |
Numbers worth knowing
- 50–60
- Questions on the paper. Pace for 60.
- 60 min
- Total time — about one minute per question.
- ~25%
- Share that are "Select TWO", with no partial credit.
- $125
- Exam fee in USD, one attempt.
- 2 years
- Validity. Recertify by retaking.
- 30%
- Core ML and AI Knowledge — the largest domain.
- 2 bytes
- Per parameter at 16-bit precision, for memory estimates.
- 4 functions
- NIST AI RMF: Govern, Map, Measure, Manage.
Acronyms
The fastest-decaying knowledge under exam pressure.
- RAG
- Retrieval-Augmented Generation
- LoRA
- Low-Rank Adaptation
- PEFT
- Parameter-Efficient Fine-Tuning
- RLHF
- Reinforcement Learning from Human Feedback
- KV cache
- Key-Value cache
- BPE
- Byte-Pair Encoding
- BLEU
- Bilingual Evaluation Understudy
- ROUGE
- Recall-Oriented Understudy for Gisting Evaluation
- NIST AI RMF
- NIST AI Risk Management Framework
- QAT
- Quantization-Aware Training
- MLM
- Masked Language Modelling
- GLUE
- General Language Understanding Evaluation
If you only read four things
- Retrieval-Augmented Generation for Knowledge-Intensive NLP TasksLewis et al., 2020 — the RAG paper, and the most valuable single read for this exam
- Attention Is All You NeedVaswani et al., 2017 — the transformer
- The Illustrated TransformerJay Alammar — read alongside the paper above
- LoRA: Low-Rank Adaptation of Large Language ModelsHu et al., 2021
Ready for a timed run?
Sit a full-length mock at exam pace, or drill the domain you are weakest in.