Nutanix NCP-AI study plan
Nutanix NCP-AI · Give each week a clear focus.
8 weeks · About 6 hours per week
Nutanix AI Stack Architecture and Components
Map each layer of the Nutanix AI stack (infrastructure, Prism Central management plane, AHV hypervisor, AOS distributed storage, NKE/Kubernetes layer, and NVIDIA NIM/NVIDIA AI Enterprise integration) to the function it performs for an AI workload
Describe how AOS distributes data and metadata across CVMs and explain why that architecture matters for dataset ingest and model checkpoint throughput
Identify which stack components are licensed or entitlement-gated versus included, and list the questions to ask before sizing a cluster for GPU workloads
Sketch a reference topology for a small GPT-in-a-Box style deployment, labeling where GPU nodes, storage-only nodes, and the Prism management plane sit
Your checkpoint
A one-page annotated architecture diagram of the Nutanix AI stack with each component labeled by function and a short note on its role in an AI training or inference flow
AI Workload Infrastructure and GPU Virtualization
Explain how AHV exposes physical GPUs to VMs, including passthrough versus virtual GPU options and the operational tradeoffs of each
Describe how GPU resources are requested, scheduled, and isolated in a Kubernetes context using device plugins and node labels
Calculate a rough GPU memory budget for a given model size, precision, and batch size, showing the arithmetic
List the host-level prerequisites (driver version, firmware, NUMA alignment, PCIe topology) that must be satisfied before a GPU node is usable
Your checkpoint
A worked GPU sizing worksheet for two scenarios (one 7B-parameter inference model and one multi-GPU fine-tuning job) with memory, node count, and storage throughput estimates
AI Workload Infrastructure and GPU Virtualization & Nutanix GPT-in-a-Box and AI-Ready Deployment
Walk through the end-to-end GPT-in-a-Box deployment sequence from cluster validation to a serving endpoint, noting the decision point at each step
Configure and validate a model-serving endpoint for an open-weight model and confirm it responds to a test prompt
Compare the storage profiles appropriate for model weights, training datasets, and checkpoint writes, and justify each choice
Diagnose three common deployment failures (insufficient GPU memory, image pull failure, storage latency on checkpoint writes) and state the first diagnostic command or Prism view to check
Your checkpoint
A reproducible deployment runbook for GPT-in-a-Box that a colleague could follow without prior context, including validation checks and rollback steps
Data Services for AI: Storage, Objects, and Databases
Differentiate Nutanix Files, Volumes, and Objects by access protocol and match each to an AI data pattern (shared training data, block-backed model cache, S3-compatible artifact store)
Explain how Nutanix Database Service provisions and manages a database instance for AI metadata or feature stores, and name the lifecycle operations it covers
Describe data protection choices for large datasets, including snapshot and replication behavior and their cost in capacity
Plan a data layout that separates hot training data from cold archives and states the tiering rule for each
Your checkpoint
A data services design document mapping four named AI data patterns to specific Nutanix storage services, with a protection and tiering decision recorded for each
Nutanix Kubernetes Platform and Containerized AI Workloads
Provision a Nutanix Kubernetes Platform cluster and deploy a containerized inference workload onto it
Explain how NKE integrates with the underlying Nutanix infrastructure for networking, storage classes, and node scaling
Configure persistent volume claims backed by Nutanix storage for a stateful AI component such as a vector database or model registry
Describe how to isolate multiple AI project namespaces with quotas, resource limits, and network policy
Your checkpoint
A running NKE cluster with one inference deployment and one stateful workload, plus a short write-up of the manifests used and why each storage class was chosen
Nutanix Kubernetes Platform and Containerized AI Workloads & Operations, Monitoring, and Lifecycle Management for AI Environments
Set up monitoring for GPU utilization, memory pressure, storage latency, and pod restarts, and identify which Prism and Kubernetes signals feed each
Define alert thresholds that would catch a degraded inference service before users notice
Describe the lifecycle operations for an AI environment: cluster upgrades, GPU driver updates, model version rollover, and decommissioning
Explain how to validate a cluster upgrade in a non-production namespace before rolling it to serving workloads
Your checkpoint
A monitoring and alerting plan listing at least six signals, their source, a threshold, and the runbook action for each
Operations, Monitoring, and Lifecycle Management for AI Environments & Security, Governance, and Multitenancy for AI
Explain the identity, role-based access, and network isolation controls that keep one AI tenant's data and compute separate from another's
Describe how to govern model artifacts and datasets so that access is auditable and versions are traceable
Draft the security review questions to ask before connecting an AI workload to production data sources
Identify the logging sources that would support a post-incident review of unauthorized access to a model endpoint
Your checkpoint
A multitenancy and governance checklist covering identity, network isolation, artifact access, and audit logging, with the owner named for each control
Nutanix AI Stack Architecture and Components & AI Workload Infrastructure and GPU Virtualization & Nutanix GPT-in-a-Box and AI-Ready Deployment & Data Services for AI: Storage, Objects, and Databases & Nutanix Kubernetes Platform and Containerized AI Workloads & Operations, Monitoring, and Lifecycle Management for AI Environments & Security, Governance, and Multitenancy for AI
Re-derive the architecture and GPU sizing answers from memory and compare against your week 1 and week 2 deliverables to find gaps
Resolve every flagged or uncertain item from earlier weeks by returning to primary documentation
Complete a timed full-length practice set and categorize each miss by domain
Write a one-page summary of the ten facts you were least confident about and the corrected version of each
Your checkpoint
A scored practice exam with a per-domain error breakdown and a corrected fact sheet covering every missed item
Milestones
Week 2: You can explain GPU passthrough versus virtual GPU and produce a defensible GPU memory estimate for a named model without consulting notes
Week 4: You can map any described AI data pattern to a specific Nutanix storage service and justify the choice in two sentences
Week 6: Your NKE cluster runs both a stateless inference deployment and a stateful workload with persistent storage, and your monitoring plan catches a simulated GPU memory saturation
Week 8: Practice score is stable across two attempts and no domain accounts for more than a quarter of total misses