Nutanix NCP-AI flashcards
Nutanix NCP-AI · Build recall with a focused review session.
32 cards. Read the term and the definition here. The deck at the top of this page is optional.
- AOS Distributed Storage Fabric
- Nutanix software-defined storage layer that pools local disks across cluster nodes, striping and replicating (RF2/RF3) data so AI datasets and model files tolerate node failure without external SAN.
- Prism Central vs Prism Element
- Prism Element manages a single Nutanix cluster; Prism Central aggregates many clusters for unified inventory, RBAC, Projects, and lifecycle management of AI environments.
- GPT-in-a-Box definition
- Nutanix validated reference architecture combining NCI, NKP, GPU nodes, Nutanix Objects, and a curated inference/MLOps stack such as NVIDIA NIM and KServe, deployed via Helm and manifests.
- Nutanix Objects role in AI
- S3-compatible object storage running on the Nutanix cluster, used as the dataset store, model registry backing store, and checkpoint target for training and inference workloads.
- GPU passthrough characteristics
- Assigns an entire physical GPU to one VM with near-native performance and full NVLink support; required for most multi-GPU training, but gives no sharing across VMs.
- NVIDIA vGPU time-slicing
- Partitions a physical data-center GPU into vGPU profiles shared by multiple VMs; suited to inference density and developer desktops, not to NVLink-scale distributed training.
- NVIDIA MIG isolation
- Multi-Instance GPU divides an A100 or H100 into hardware-isolated slices with dedicated memory and SMs, each passable to a separate VM for strict multitenant separation.
- LLM VRAM sizing rule
- Approximate weights memory equals parameter count times bytes per parameter: FP16 about 2 bytes, INT8 about 1, INT4 about 0.5; add KV cache for context.
- Tensor versus pipeline parallelism
- Tensor parallelism splits layers across GPUs inside one node over NVLink; pipeline parallelism splits stages across nodes over the network, which is slower and harder to balance.
- NVIDIA GPU Operator purpose
- Kubernetes operator that installs and manages GPU drivers, container runtime hooks, device plugin, and DCGM monitoring on NKP GPU nodes, removing manual guest driver installs.
- KServe InferenceService
- Kubernetes custom resource for model serving that provides autoscaling including scale-to-zero, canary traffic splitting, and standardized model endpoints for GPT-in-a-Box deployments.
- GPT-in-a-Box deployment order
- Install NCI, enable GPU passthrough or vGPU, deploy NKP, create GPU node pools, install GPU Operator, then deploy the AI stack and load models from Objects.
- Air-gapped AI stack prerequisites
- Offline GPT-in-a-Box installs need a local container registry mirror and pre-pulled large images, planned in the hardware bill of materials before deployment begins.
- Storage container settings
- AOS policy unit defining replication factor, compression, deduplication, and erasure coding; dedupe helps repetitive datasets but not already-compressed model checkpoints.
- Nutanix Files for datasets
- SMB and NFS file service on the Nutanix cluster, commonly mounted over NFS for shared training datasets and exposed to Kubernetes as ReadWriteMany persistent volumes.
- NDB thin clone
- Nutanix Database Service creates near-instant space-efficient clones of Postgres, MySQL, MongoDB, or SQL Server instances, useful for per-experiment metadata and feature databases.
- Nutanix CSI driver
- Container Storage Interface driver that surfaces AOS volumes and Files shares as Kubernetes persistent volume claims for pods running on Nutanix Kubernetes Platform.
- Objects versioning and WORM
- Bucket-level object versioning preserves prior artifact versions, while write-once-read-many retention makes model and audit data immutable for governance requirements.
- NKP GPU resource request
- Pods request GPUs through resources.limits nvidia.com/gpu; the NVIDIA device plugin advertises the resource, and extended resources require limits equal to requests.
- NKP node pools
- Groups of worker nodes with identical configuration; GPU node pools define GPU model and count, and workloads use node selectors or affinity to target them.
- Scale-to-zero inference
- KServe with Knative scales idle model endpoints to zero pods and back on demand, controlling cost of expensive GPU nodes in Nutanix Kubernetes Platform clusters.
- NKP namespace multitenancy
- Team isolation on Nutanix Kubernetes Platform using namespaces with RBAC, ResourceQuota including nvidia.com/gpu, NetworkPolicy, and LimitRange for predictable sharing.
- DCGM GPU telemetry
- NVIDIA Data Center GPU Manager exporter exposes utilization, memory, temperature, ECC, and XID errors to Prometheus for dashboards and alerting on AI clusters.
- AI inference SLO metrics
- Monitor GPU utilization and memory, request latency percentiles, tokens per second, queue depth, and error rate per model endpoint to detect degradation before users do.
- LCM upgrade ordering
- Life Cycle Manager updates firmware and NCI software; verify GPU driver and vGPU manager against the compatibility matrix before upgrading AI clusters.
- GPU maintenance draining
- Cordon and drain GPU nodes before host maintenance so training jobs checkpoint to Objects or NFS and reschedule onto healthy nodes without losing progress.
- Prism Central Projects
- RBAC scoping construct that isolates VMs, networks, images, and quotas per tenant, giving each AI team a bounded environment with role-based access.
- Flow Network Security
- Nutanix microsegmentation that applies distributed firewall policy per VM using categories rather than IP addresses, isolating AI tenants and restricting inference pod egress.
- Objects bucket policies
- S3-style access controls with scoped access keys and pre-signed URLs that grant time-limited access to model artifacts and datasets without exposing long-term credentials.
- Data-at-rest encryption options
- Nutanix supports self-encrypting drives with external key management or software-based cluster encryption; the key manager choice determines compliance and audit posture.
- Model lineage governance
- Recording which dataset version, code commit, and hyperparameters produced each model artifact in Objects, with a GitOps review gate before production promotion.
- Inference gateway logging
- Prompt and response logging with PII redaction placed at the inference gateway rather than inside the model, supporting audit while limiting sensitive data exposure.