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

Free NCA-AMAD NVIDIA AI Enterprise and Blueprints Overview Practice Questions

The NVIDIA AI Enterprise and Blueprints Overview domain is 10% of the NVIDIA NCA-AMAD exam and covers NVIDIA NIM, NVIDIA AI Enterprise, Blueprints, Kubernetes deployment, hybrid and self-hosted deployment and the AI Factory. 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 NVIDIA AI Enterprise and Blueprints Overview 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 NVIDIA AI Enterprise and Blueprints Overview

NCA-AMAD NVIDIA AI Enterprise and Blueprints Overview sample questions with answers

NVIDIA AI Enterprise and Blueprints Overview

Sample question 1

In the NVIDIA NIM Operator for Kubernetes, what is the job of a NIMCache custom resource?

  1. It adds or removes NIM pods as the GPU load on the cluster changes
  2. It downloads model files ahead of time and keeps them on a persistent volume
  3. It packages the model weights into a new container image for each release
  4. It spreads incoming inference requests evenly across the running NIM pods
Show correct answer and explanation

Correct answer: B. It downloads model files ahead of time and keeps them on a persistent volume

B is correct because a NIMCache pulls the model profiles from NGC or another source and persists them on a PVC, so NIM pods can start from the cached copy instead of downloading at startup. A is incorrect because scaling pods is the job of a horizontal pod autoscaler, which the NIMService resource can create. C is incorrect because the cache stores model files on a volume; it does not build container images. D is incorrect because request distribution is handled by the Kubernetes service or an ingress, not by the cache resource.

NVIDIA AI Enterprise and Blueprints Overview

Sample question 2

NVIDIA Blueprints and NIM microservices are often delivered with a Helm chart. What does a Helm chart provide?

  1. A container image that holds the model weights and the inference engine together
  2. A GPU driver package that must be installed on every node before serving
  3. A benchmark report that shows the latency and throughput of the model
  4. A packaged, configurable set of Kubernetes manifests that installs with one command
Show correct answer and explanation

Correct answer: D. A packaged, configurable set of Kubernetes manifests that installs with one command

D is correct because Helm bundles the Deployments, Services and other Kubernetes objects an application needs, with values the team can override, so the whole stack installs and upgrades as one release. A is incorrect because the container image is what the chart deploys; the chart itself is a set of templated Kubernetes manifests. B is incorrect because GPU drivers are managed separately, for example by the GPU Operator; a Helm chart describes application resources. C is incorrect because performance figures come from tools such as genai-perf, not from the deployment package.

NVIDIA AI Enterprise and Blueprints Overview

Sample question 3

A platform team on Kubernetes already serves LLM NIM model files from a shared cache, yet new pods still take a long time to become ready during scale-out because each one builds its TensorRT-LLM engine at startup. Which change BEST shortens pod startup?

  1. Prebuild the engines on the target GPU type before the pods are deployed
  2. Raise minReplicas so that more pods have finished building before a spike
  3. Lengthen the readiness probe timeout so that slow pods are not restarted
  4. Move the shared model cache to faster storage so that the files load sooner
Show correct answer and explanation

Correct answer: A. Prebuild the engines on the target GPU type before the pods are deployed

A is correct because the NIM Operator can generate optimized TensorRT-LLM engines from the NIM's model profiles ahead of time on the desired GPUs, so new pods load a ready engine instead of building one, which makes startup and autoscaling faster and more predictable. B is incorrect because this keeps idle GPUs running all the time, and every pod added during a spike still builds its own engine. C is incorrect because a longer timeout only tolerates the delay; each pod still builds its engine before it can serve. D is incorrect because the files are already cached; the slow step is the compute-heavy engine build that every pod repeats.

NVIDIA AI Enterprise and Blueprints Overview

Sample question 4

An engineer applies this resource to a cluster that serves an LLM NIM:

kind: HorizontalPodAutoscaler
spec:
  scaleTargetRef:
    name: meta-llama3-8b
  minReplicas: 1
  maxReplicas: 10
  metrics:
  - type: Pods
    pods:
      metric:
        name: gpu_cache_usage_perc
      target:
        type: AverageValue
        averageValue: 100m

What signal will this autoscaler use to add or remove NIM pods?

  1. The CPU utilization of the nodes that host the NIM pods
  2. The GPU compute utilization that DCGM reports for each node
  3. The p95 request latency measured at the cluster ingress
  4. The average KV cache utilization reported by the NIM pods
Show correct answer and explanation

Correct answer: D. The average KV cache utilization reported by the NIM pods

D is correct because gpu_cache_usage_perc is the NIM metric for the percent of KV cache in use, and the Pods metric type with an AverageValue target scales on its average across the pods. A is incorrect because CPU is a standard resource metric, while this autoscaler reads a custom per-pod metric. B is incorrect because the metric named is the NIM's KV cache usage, not a DCGM compute counter. C is incorrect because no latency metric is referenced; latency is a possible extra metric but not the one configured here.

NVIDIA AI Enterprise and Blueprints Overview

Sample question 5

A team wants its HPA to scale an LLM NIM on the gpu_cache_usage_perc metric. Prometheus already scrapes the NIM's metrics endpoint, but the HPA reports the metric as unavailable. Which component should the team install or fix NEXT?

  1. A metrics server that reports pod CPU and memory to the resource metrics API
  2. An adapter that exposes the Prometheus metric through the custom metrics API
  3. A genai-perf load test that drives traffic to generate the KV cache metric
  4. A Grafana dashboard that charts the KV cache metric for the operations team
Show correct answer and explanation

Correct answer: B. An adapter that exposes the Prometheus metric through the custom metrics API

B is correct because the HPA reads custom metrics from the Kubernetes custom metrics API, and a Prometheus adapter is the component that serves Prometheus data there, pointed at the right Prometheus service. A is incorrect because a metrics server scrapes CPU and memory from kubelets; it does not expose custom Prometheus metrics such as KV cache usage. C is incorrect because genai-perf only generates benchmark traffic; it does not make metrics available to the autoscaler. D is incorrect because Grafana visualizes metrics for people; the HPA cannot read a dashboard.

NVIDIA AI Enterprise and Blueprints Overview

Sample question 6

A load test ends and the KV cache metric drops well below its target, but the HPA keeps three NIM replicas running:

$ kubectl get hpa -n llm
NAME           REFERENCE                  TARGETS    MINPODS  MAXPODS  REPLICAS
gpu-hpa-cache  Deployment/meta-llama3-8b  293m/100m  1        10       3
gpu-hpa-cache  Deployment/meta-llama3-8b  7m/100m    1        10       3

Which explanation is MOST likely?

  1. The custom metric stopped updating, so the HPA froze the replica count
  2. The maxReplicas value of 10 forces the HPA to keep extra pods running
  3. The NIM pods refuse to stop while their KV cache still holds data
  4. The HPA waits out a downscale stabilization window before it removes pods
Show correct answer and explanation

Correct answer: D. The HPA waits out a downscale stabilization window before it removes pods

D is correct because Kubernetes delays scale-down through a stabilization window so that rapidly changing metrics do not cause replicas to flap; the extra pods are removed gradually once the window passes. A is incorrect because the TARGETS column shows the metric updating to 7m, so the autoscaler is receiving fresh values. B is incorrect because maxReplicas is an upper limit; it never forces pods to stay when the metric is low. C is incorrect because pods do not block termination because of cache contents; the delay comes from the autoscaler's own policy.

NVIDIA AI Enterprise and Blueprints Overview

Sample question 7

A RAG application runs as separate Kubernetes services: a frontend, an embedding NIM, a reranking NIM and an LLM NIM. Security now requires encrypted, mutually authenticated traffic between these services and per-service traffic policies, without changing application code. What should the team add?

  1. An ingress controller that terminates TLS for requests coming from outside
  2. A horizontal pod autoscaler for each NIM that scales on KV cache usage
  3. A service mesh that places proxies beside each service to handle mTLS and policy
  4. A NIMCache resource that stores every model on encrypted shared storage
Show correct answer and explanation

Correct answer: C. A service mesh that places proxies beside each service to handle mTLS and policy

C is correct because a service mesh intercepts service-to-service traffic with sidecar or node proxies, so it can add mutual TLS, retries and traffic policies to existing microservices without code changes. A is incorrect because an ingress handles traffic entering the cluster; it does not secure the calls the services make to each other. B is incorrect because autoscaling changes replica counts and does nothing for encryption or traffic policy. D is incorrect because encrypting cached model files protects data at rest, not the requests flowing between services.

NVIDIA AI Enterprise and Blueprints Overview

Sample question 8

A team autoscales an LLM NIM on node CPU utilization with a 70% target. During a traffic spike, users see latency climb sharply, yet CPU stays near 20% and no pods are added. Which change is MOST likely to make autoscaling follow the real load?

  1. Scale on host memory usage because model weights live in memory
  2. Lower the CPU target to 15% so that the autoscaler reacts much earlier
  3. Raise maxReplicas so that the autoscaler is allowed to add more pods
  4. Scale on a NIM-specific metric such as KV cache utilization instead of CPU
Show correct answer and explanation

Correct answer: D. Scale on a NIM-specific metric such as KV cache utilization instead of CPU

D is correct because LLM inference load sits on the GPU, so CPU barely moves; NIM exposes metrics such as gpu_cache_usage_perc that rise with concurrency, which makes them a better scaling signal. A is incorrect because model weights and KV cache sit in GPU memory, and host memory use stays flat as traffic grows. B is incorrect because CPU does not track GPU-bound inference load, so a lower target would scale on noise rather than on real demand. C is incorrect because the autoscaler is not hitting its ceiling; it never asks for more pods because the CPU signal stays low.

NVIDIA AI Enterprise and Blueprints Overview

Sample question 9

A production cluster runs many replicas of one LLM NIM, and the scheduler places each new pod on whichever of several GPU nodes has free capacity, often a node that has never run the model. Each new pod downloads about 50 GB of model files from NGC, so scale-out takes many minutes. Which change BEST shortens pod startup on every node?

  1. Use the Local Path Provisioner so that each node keeps its own cache volume
  2. Mount an emptyDir volume at the NIM cache path inside each pod
  3. Raise minReplicas so that more pods are already running before the spike
  4. Cache the model once in a NIMCache on storage that all nodes share
Show correct answer and explanation

Correct answer: D. Cache the model once in a NIMCache on storage that all nodes share

D is correct because a NIMCache on a network-backed, ReadWriteMany volume lets every replica on any node start from one cached copy, which is what NVIDIA recommends for multi-pod production deployments. A is incorrect because a node-local volume suits single-node development; a pod scheduled on a node that has not cached the model still downloads it. B is incorrect because emptyDir is created fresh for each pod and deleted with it, so every new replica still downloads the model. C is incorrect because this pays for idle GPUs all the time and still leaves each added pod with a slow download.

NVIDIA AI Enterprise and Blueprints Overview

Sample question 10

Engineers deploy NIM services to staging and production by running kubectl edit by hand. The two clusters have drifted apart, and auditors cannot tell who changed what. Which approach BEST makes deployments repeatable and auditable?

  1. Copy the production namespace into staging every night with a sync script
  2. Turn on Kubernetes audit logging and keep making changes with kubectl edit
  3. Restrict kubectl edit to one release engineer who keeps a written change log
  4. Store the manifests and Helm values in Git and apply them through a pipeline
Show correct answer and explanation

Correct answer: D. Store the manifests and Helm values in Git and apply them through a pipeline

D is correct because declarative manifests in version control, applied automatically, give the same result in every cluster and leave a reviewed history of every change, which is the core of infrastructure as code. A is incorrect because nightly copies spread whatever drift exists and record nothing about who made each change or why. B is incorrect because audit logs record who changed what, but hand edits still let the clusters drift apart, so deployments are not repeatable. C is incorrect because one person typing changes by hand can still make mistakes and drift, and a manual log is not a reliable audit trail.

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.