Nutanix NCP-AI common mistakes
Nutanix NCP-AI · Understand the traps and how to avoid them.
Treating GPT-in-a-Box as a single appliance and describing it as one bundled VM, rather than as a reference architecture that combines Nutanix Cloud Infrastructure, NKE/Nutanix Kubernetes Platform, Files or Objects storage, and a validated model-serving stack.
Marketing summaries compress the offering into a product name, and candidates who only skim the landing page never internalize the layered architecture, so they misplace components when asked which layer provides GPU scheduling versus model serving versus persistent storage.
What to do instead
Build a one-page layer map: infrastructure (AHV clusters, GPU nodes), Kubernetes layer (NKP/NKE), storage layer (Files for model artifacts, Objects for datasets and checkpoints), and serving layer (NVIDIA Triton, vLLM, or similar inference runtimes). For each exam scenario, identify which layer owns the responsibility before answering.
Selecting vGPU or MIG partitioning for a workload that requires the full framebuffer or all SMs, or assuming GPU passthrough and vGPU have identical live-migration behaviour.
Candidates memorize that partitioning exists but not the tradeoffs: MIG slices SMs and memory with no vMotion support in typical configurations, while passthrough devices also block standard live migration; vGPU profiles carry their own licensing and capacity constraints.
What to do instead
Learn a decision table keyed on workload: large-model training wants full GPUs with NVLink locality, many small inference tenants want MIG or time-sliced vGPU, and any workload needing vMotion or HA placement rules must use supported vGPU profiles. Practice mapping three sample workloads to a GPU allocation choice and justify it in one sentence.
Assuming any cluster with GPUs can host AI workloads without checking AHV host sizing, GPU compatibility matrix entries, power and cooling headroom, and whether the GPU model is on the supported list for the installed AOS and NKP versions.
The exam tests planning discipline, but candidates often study features instead of prerequisites, so they miss that GPU enablement depends on host firmware, driver and VIB versions, and validated node models rather than just physical card presence.
What to do instead
Memorize the pre-deployment checklist as a sequence: validate node and GPU on the compatibility matrix, confirm AOS and NKP version support, install correct NVIDIA drivers and VIBs, enable GPU passthrough or vGPU in AHV, then verify with nvidia-smi inside a guest. Rehearse this order because questions often ask what to do first.
Confusing Nutanix Files, Objects, and Volumes when choosing storage for model artifacts, training datasets, and checkpoint writes, and defaulting to Volumes for everything.
All three are presented as data services, and candidates who lack hands-on exposure do not distinguish SMB/NFS semantics from S3 object semantics or block-level iSCSI, so they pick the familiar-sounding option instead of matching access pattern and concurrency needs.
What to do instead
Anchor each service to an access protocol and use case: Files for shared NFS/SMB model repositories and multi-reader datasets, Objects for S3-compatible immutable datasets and large-scale checkpoint or artifact retention, Volumes for block workloads needing raw device semantics. For each practice question, state the protocol first, then choose the service.
Answering Kubernetes questions with generic upstream Kubernetes behaviour and ignoring Nutanix-specific constructs such as NKP cluster lifecycle, CSI storage provisioning, and how GPU device plugins are exposed to pods.
Candidates with CKA-style backgrounds rely on prior knowledge, but the exam asks how NKP integrates with Prism Central, how the Nutanix CSI driver presents storage classes, and how NVIDIA device plugins advertise GPU resources, which differ from vanilla distributions.
What to do instead
Study the NKP-specific path: cluster creation and upgrade through the Nutanix management plane, the Nutanix CSI driver and its StorageClasses, and the NVIDIA GPU Operator or device plugin flow that makes nvidia.com/gpu a schedulable resource. Be able to trace a pod from PVC to volume and from GPU request to assigned device.
Ignoring resource quotas, namespaces, and project-level isolation when asked to design a multitenant AI platform, answering only with network segmentation or RBAC.
Multitenancy questions feel like security questions, so candidates reach for the most familiar control and neglect the compute and storage dimensions, missing that GPU quotas, namespace-scoped storage classes, and Prism Central projects are the primary isolation mechanisms.
What to do instead
Answer multitenancy questions with four axes every time: identity and RBAC, network segmentation and flow policies, compute quotas including GPU limits per namespace or project, and storage isolation through separate projects, shares, or buckets. State which axis the question targets before choosing.
Describing monitoring only in terms of VM-level CPU and memory alerts, omitting GPU telemetry such as utilization, memory pressure, ECC errors, temperature, and XID errors, and omitting inference metrics like latency and queue depth.
Traditional virtualization monitoring habits dominate, and candidates do not realize that GPU health and model-serving SLOs require separate exporters and dashboards, so they cannot answer questions about detecting a degraded GPU or a saturated inference endpoint.
What to do instead
Memorize two telemetry planes: infrastructure (Prism Central alerts, NCC health checks, GPU metrics via DCGM exporter) and application (request latency, throughput, error rate, queue depth from the serving framework). For any monitoring question, name the metric, the collector, and the alert threshold rationale.
Treating lifecycle management as a single upgrade action, rather than sequencing AOS, AHV, firmware, NKP, GPU drivers, and model runtime upgrades with compatibility checks between each step.
Candidates think in terms of one product's upgrade button, but AI environments have a dependency chain where upgrading the platform ahead of validated GPU drivers or the Kubernetes version can break workloads, so order matters more than the individual steps.
What to do instead
Learn the upgrade order and the reason for it: firmware and compatibility validation first, then AOS/AHV, then NKP and CSI, then GPU operator and drivers, then model runtime images. Practice explaining why a driver upgrade after a platform upgrade avoids unsupported combinations.
Answering security questions with perimeter controls only, missing data protection for AI assets such as encryption of model artifacts and datasets at rest, bucket and share access policies, secrets management for inference endpoints, and audit trails.
AI security feels like a network problem, but the exam emphasizes data-centric governance, and candidates without governance experience default to firewalls and forget that model weights and training data are the crown jewels requiring their own controls.
What to do instead
Structure answers around the data lifecycle: classify assets, encrypt at rest and in transit, restrict access through least-privilege policies on Files shares and Objects buckets, manage credentials in a secrets store, and enable auditing. Be ready to name which Nutanix capability enforces each control.
Guessing at GPU capacity and sizing numbers, such as how many concurrent inference requests a card supports or how much memory a model needs, instead of reasoning from model size, precision, and context length.
Candidates memorize figures from a single blog post and then apply them to scenarios with different models or precisions, forgetting that memory footprint scales with parameters times bytes per parameter plus KV cache and overhead.
What to do instead
Learn the estimation method rather than the number: parameters times bytes per parameter for the chosen precision, plus activation and KV cache overhead, plus headroom, then compare with GPU memory and multiply by node count. Show the arithmetic in your reasoning so a wrong constant does not sink the answer.