NCA-GENM — NVIDIA-Certified Associate

Prep for NCA-GENM

300 practice questions on multimodal generative AI — images, audio, video and the pipelines that carry them — with an explanation written for every one of the four options. Entirely offline.

Coming soon: App Store Coming soon: Google Play

• Exam NCA-GENM, Generative AI Multimodal (NVIDIA-Certified Associate)
• Questions 300, split into 6 practice sets of 50
• Explanations 1,200 — one per answer option
• Price USD 4.99, one-time. No in-app purchases
• Network None. The app never connects
• Platforms iOS and Android

The exam this app is for

NCA-GENM is the associate-level exam for generative AI once text stops being the only modality. Diffusion models and how they are conditioned, encoders for image, audio and video, embeddings from two models that have to be made to line up, and the practical handling that comes with media: sample rates, variable-length clips, padding and masks, frames from a video.

It is the most hands-on of the three associate exams here. A surprising share of the questions come from the data side — captions longer than the model accepts, clips recorded with the gain too high, a train/test split that leaks because the clips were cut from the same source video — and from the evaluation side, where a two-percent lead after one day is not a result. If you have ever shipped a media pipeline, the scenarios will feel familiar.

What the 300 questions cover

Seven domains, with the app's own distribution below. The shares are an editorial choice for practice, not an official blueprint.

• Experimentation72 questions · 24%

  Comparing two pipelines honestly: what an early A/B lead is worth, splits that do not leak across clips from one source, controlling the variables when two prompt phrasings are compared, and the pitfalls of grading a small model with a large one.

• Core Machine Learning and AI60 questions · 20%

  How diffusion actually works, inpainting a masked region, conditioning a generation on an edge map when no wording gets you there, what raising decoder temperature does to a caption, and adapting a model per customer without storing a full checkpoint each time.

• Multimodal Data48 questions · 16%

  Resampling audio to the rate a model was trained on, batching clips of different lengths without letting the padding leak into the output, representing a video clip for a transformer, and what has to be true before two separately trained encoders can be fused.

• Software Development48 questions · 16%

  Serving a preprocessing step, a vision encoder and a decoder as one server-side pipeline, keeping similarity search fast over a large catalogue, choosing batch processing when nobody is waiting, and monitoring whether output quality is still holding after deployment.

• Data Analysis and Visualization30 questions · 10%

  Finding the problems in a dataset before fine-tuning on it: captions past the input limit, clipped audio from a bad recording gain, clustering two million unlabelled images to learn what categories exist, and what t-SNE shows that PCA does not.

• Performance Optimization30 questions · 10%

  Where the time actually goes: host-to-device copies stalling the loop, launch overhead dominating at batch size one, what building a TensorRT engine changes, and distilling a model that is accurate but too slow to ship.

• Trustworthy AI12 questions · 4%

  Layered defences on an image-generation service, watermarking published output, consent and provenance when face data was scraped, and what a Shapley-value attribution does and does not tell a reviewer about one prediction.

How you study with it

• Six sets of 50, each mixing all seven domains, so the small domains still appear in every rehearsal instead of being skipped.
• Exam mode runs the set against a clock with no feedback until the end. Practice mode explains each answer on the spot.
• Domain scoring after every set. On this exam the split is usually clean: people who work with models score on the core and optimisation domains and lose points on data handling, or exactly the reverse.
• Wrong answers and bookmarks become their own review runs, which is the fastest way to clear a domain you keep missing.
• Four explanations per question, because in this bank the distractors are mostly techniques that are real and simply solve a different problem — upscaling instead of conditioning, a bigger batch instead of a faster kernel.

Questions people ask

Is this only about image generation?

No. Images are the largest single thread, but audio and video carry real weight — sample rates, padding and masking, frame sampling, ASR data problems — and so does the text side of a vision-language model. "Multimodal" in this exam means the joins between modalities as much as the modalities themselves.

Do I need a GPU or any software installed?

Not for the app. It is a question bank on a phone. The questions do assume you have some idea of what a training loop and an inference server look like, and the explanations fill in the rest.

I already passed NCA-GENL. Does that cover this one?

Partly. The shared ground is the general machine-learning and evaluation material. Almost everything specific — diffusion, encoders, media preprocessing, the optimisation domain — is new, which is why Prep for NCA-GENL and this app are separate banks with separate questions.

How heavy is the maths?

Light. You read a result, compare two options and decide what is sound. There is no derivation of a diffusion objective, and no calculator is needed.

Will it keep working with no signal?

Yes. The whole bank is in the download, the app makes no network requests at all, and all progress is stored on the device. See the privacy policy.

Where do the questions come from?

They are written with AI and organised against the published NCA-GENM domains. They are original practice material, not real exam questions, and no exam question is reproduced. Spot a mistake? Write to us and it is fixed for everyone in the next update.

← All four apps

Disclaimer. Prep for NCA-GENM is an independent study aid. It is not affiliated with, authorized by, endorsed by, or sponsored by NVIDIA Corporation. NVIDIA, NCA-GENM and the NVIDIA-Certified Associate program name are trademarks or certification program names of NVIDIA Corporation, used here only to identify the exam this app helps you prepare for. No certification result is guaranteed.