NVIDIA NCA-AMAD exam · Official objectives · Original practice content · Last reviewed October 16, 2026

Free NCA-AMAD Sizing and Configuration Practice Questions

The Sizing and Configuration domain is 10% of the NVIDIA NCA-AMAD exam and covers GPU memory sizing, replicas and scaling, batch size and precision, Run:ai, model lifecycle, monitoring, MTTD and MTTR. Review each answer and explanation as you practice.

The NCA-AMAD exam has 50-60 multiple-choice questions plus a hands-on lab in 1 hour. It costs $125, results are pass or fail, and the certification is valid for 2 years.

Take the 10-question NCA-AMAD quick-start test · NCA-AMAD cheat sheet · Create a free account

What Sizing and Configuration covers on the NCA-AMAD exam

This domain is 10% of the NCA-AMAD exam (about 6 questions in a 60-question mock). NVIDIA's objectives include:

How to study Sizing and Configuration

NCA-AMAD Sizing and Configuration sample questions with answers

Sizing and Configuration

Sample question 1

How does a data flywheel help keep a deployed model from drifting out of date?

  1. It swaps in the newest public model release on a fixed monthly schedule
  2. It freezes the model weights so that behavior never changes after release
  3. It raises the sampling temperature as incoming queries become more varied
  4. It evaluates real usage data and triggers retraining when quality slips
Show correct answer and explanation

Correct answer: D. It evaluates real usage data and triggers retraining when quality slips

D is correct because the flywheel captures real-world usage, evaluates model performance and triggers retraining or fine-tuning as needed, which keeps the model aligned with current users and context. A is incorrect because a calendar swap ignores whether quality actually dropped and skips evaluation on the team's own data. B is incorrect because frozen weights are exactly what lets a model drift out of date as data and user behavior change. C is incorrect because temperature only changes output randomness; it does not teach the model anything new.

Sizing and Configuration

Sample question 2

In NVIDIA NeMo microservices, which component stores the files behind datasets and models and supports versioning them?

  1. NeMo Retriever
  2. NeMo Guardrails
  3. NeMo Evaluator
  4. NeMo Datastore
Show correct answer and explanation

Correct answer: D. NeMo Datastore

D is correct because NeMo Datastore handles the actual files behind datasets and models, including uploading, downloading and versioning, while NeMo Entity Store tracks the entities themselves. A is incorrect because NeMo Retriever extracts, embeds and reranks content for retrieval; it is not a dataset and model file store. B is incorrect because NeMo Guardrails enforces safety and topic policies on inputs and outputs; it does not store or version files. C is incorrect because NeMo Evaluator runs benchmarks and custom evaluations; it reads datasets but does not manage their versions.

Sizing and Configuration

Sample question 3

A banking agent was fine-tuned to write SQL queries against a transactions database. The bank then moves part of its data to a new store with a different schema, and the agent keeps generating queries for the old structure. Which lifecycle action BEST addresses this?

  1. Retrain the agent's model on examples that use the new schema
  2. Roll back to the previous model version from before the move
  3. Lower the sampling temperature so the generated queries are more stable
  4. Add replicas so the agent can retry failed queries more quickly
Show correct answer and explanation

Correct answer: A. Retrain the agent's model on examples that use the new schema

A is correct because the data sources and responses the model depends on have changed, which is a form of model drift; without retraining on data that reflects the new schema, the agent keeps producing queries for the old structure. B is incorrect because the earlier version was trained on the same old schema, so it fails the same way. C is incorrect because more deterministic output still targets the old schema; temperature cannot add knowledge of the new one. D is incorrect because more capacity repeats the same wrong queries faster; the problem is model knowledge, not throughput.

Sizing and Configuration

Sample question 4

A retail support team runs the NVIDIA Data Flywheel Blueprint against its 70B tool-calling agent. A fine-tuned 1B candidate now scores close to the 70B model on the evaluation built from production logs. What is the NEXT step in the release process?

  1. Retrain the candidate from scratch so it no longer depends on the teacher
  2. Merge the candidate's LoRA weights back into the 70B model and redeploy it
  3. Review the results, then promote the candidate to replace the larger NIM
  4. Delete the production logs because the candidate no longer needs them
Show correct answer and explanation

Correct answer: C. Review the results, then promote the candidate to replace the larger NIM

C is correct because the blueprint ends each cycle with review and promotion: developers check the metrics and artifacts, then promote a top-performing candidate to production in place of the larger model. A is incorrect because the LoRA-tuned candidate already meets the bar; retraining from scratch would waste compute and discard a validated model. B is incorrect because the goal is to replace the large model with a smaller, cheaper one, not to add the small model's adapter to the 70B model. D is incorrect because the flywheel keeps using fresh production logs for later cycles; the loop does not end at one promotion.

Sizing and Configuration

Sample question 5

A manufacturing team wants a larger LoRA adapter for the next cycle of its Data Flywheel Blueprint, while a flywheel run is still in progress. The training settings live in the blueprint's configuration file. What should the team do?

  1. Send the new adapter size in the request that starts the next job
  2. Change the adapter size in the evaluation service's own settings
  3. Edit the file now so that the running job applies it next epoch
  4. Stop the services, update the configuration file, then redeploy
Show correct answer and explanation

Correct answer: D. Stop the services, update the configuration file, then redeploy

D is correct because the blueprint reads its configuration file when the system starts and keeps those settings fixed during a run, so a training change needs the services stopped, the file updated and the blueprint redeployed. A is incorrect because the job request identifies the workload to process; training settings come from the configuration file. B is incorrect because the evaluation service scores candidates; the LoRA training settings come from the configuration file and are used by the customization service. C is incorrect because the file is read only at startup; edits made during a run are not applied to it.

Sizing and Configuration

Sample question 6

A healthcare team deploys its NIM with the manifest excerpt below:

containers:
  - name: notes-llm
    image: nvcr.io/nim/meta/llama-3.1-8b-instruct:latest
    resources:
      limits:
        nvidia.com/gpu: 1

After a routine node restart, the pod pulls a newer container, and answers in the validated clinical-notes workflow change. Which practice BEST prevents this?

  1. Lower the temperature so the newer model's answers match the old ones
  2. Mount a host cache so the new container downloads faster after restarts
  3. Pin a specific, validated image tag and promote new versions deliberately
  4. Add a second GPU so the container has more memory after each restart
Show correct answer and explanation

Correct answer: C. Pin a specific, validated image tag and promote new versions deliberately

C is correct because the latest tag moves whenever a new release is published, so a restart can silently change the model or engine; pinning a tested version (as the blueprint config does with its tag field) makes each release deliberate and reproducible. A is incorrect because sampling settings cannot make a different model version behave like the validated one. B is incorrect because a cache speeds up startup but still serves whatever version the latest tag points to. D is incorrect because the change came from a different container version, not from a shortage of GPU memory.

Sizing and Configuration

Sample question 7

A finance team discovers a regression in model version 3, which was trained four months ago, and needs to rebuild that exact version to investigate. Which record makes this MOST feasible?

  1. The current production config, which lists the newest model's settings
  2. The serving dashboard showing version 3's latency and throughput
  3. A copy of the inference logs collected while version 3 was serving
  4. Lineage metadata linking the version to its datasets, parameters and job
Show correct answer and explanation

Correct answer: D. Lineage metadata linking the version to its datasets, parameters and job

D is correct because lineage tracking stores the training jobs, datasets, parameters and endpoints behind each model, so the team can find the exact artifacts and reproduce version 3. A is incorrect because the current config describes the latest version, not the one trained four months ago. B is incorrect because performance metrics describe how the model ran, not which data and settings produced it. C is incorrect because inference logs show requests and responses, not the training data and parameters needed to rebuild the model.

Sizing and Configuration

Sample question 8

An e-commerce search assistant is retrained from scratch every night, which consumes most of the team's GPU budget. Evaluation shows quality stays flat for weeks and then drops after large catalog changes. Which retraining policy BEST balances freshness and cost?

  1. Keep nightly retraining from scratch so the model is never out of date
  2. Trigger fine-tuning on recent curated data when evaluation scores drop
  3. Retrain from scratch once a quarter regardless of evaluation results
  4. Stop retraining and rely on the original model until users complain
Show correct answer and explanation

Correct answer: B. Trigger fine-tuning on recent curated data when evaluation scores drop

B is correct because a flywheel retrains or fine-tunes as needed when evaluation shows decline, and fine-tuning on the most relevant new data keeps accuracy current at a fraction of the cost of nightly full retraining. A is incorrect because nightly full retraining spends most of the budget during the weeks when quality is flat and adds nothing. C is incorrect because a fixed quarterly schedule can leave the model degraded for weeks after a large catalog change. D is incorrect because waiting for complaints lets quality fall in production and turns users into the detection system.

Sizing and Configuration

Sample question 9

A RAG assistant's answer quality dropped after a release that changed the LLM, the prompt template and the embedding model used for the document index. The team can roll back the LLM but cannot tell which prompt and index versions served it before. Which release practice BEST prevents this next time?

  1. Tag each release with its deployment date and keep the change notes in a wiki
  2. Keep only the newest version of each component to avoid confusion
  3. Version only the LLM weights, since the model is what generates each answer
  4. Version the model, prompt template and index together as one release with lineage
Show correct answer and explanation

Correct answer: D. Version the model, prompt template and index together as one release with lineage

D is correct because the assistant's behavior depends on all three components, so recording them as one versioned release with lineage lets the team reproduce or roll back the exact combination that worked. A is incorrect because a date and free-text notes do not record which exact artifacts were deployed, so the earlier combination still cannot be rebuilt reliably. B is incorrect because discarding earlier versions removes the ability to roll back to a known-good release. C is incorrect because the prompt template and the index also change the answers; versioning the weights alone is what left the team unable to restore the earlier behavior.

Sizing and Configuration

Sample question 10

A customer-support chatbot's weekly operations review shows four changes. Which one is the BEST trigger to start a retraining cycle rather than a scaling or infrastructure change?

  1. Daily request volume grew by 40% after a new product launch
  2. p95 time to first token rose from 600 ms to 1.4 s during peak hours
  3. Task accuracy on a fresh sample of recent tickets fell from 92% to 81%
  4. Average GPU utilization climbed from 55% to 85% over the week
Show correct answer and explanation

Correct answer: C. Task accuracy on a fresh sample of recent tickets fell from 92% to 81%

C is correct because a drop in accuracy on current data signals model drift, which is what retraining or fine-tuning addresses; the other signals describe capacity or load. A is incorrect because more requests call for more replicas or capacity; volume alone does not mean the model is answering worse. B is incorrect because rising latency under peak load points to capacity or queuing, which scaling addresses, not retraining. D is incorrect because higher utilization reflects more traffic or heavier requests; it says nothing about answer quality.

Keep practicing

NCA-AMAD practice test hub · Mixed quick start · NCA-GENL practice test

FlashGenius is an independent study resource and is not affiliated with or endorsed by NVIDIA. NVIDIA and NCA-AMAD are trademarks of NVIDIA Corporation. 600 original practice questions.