Nutanix NCP-AI key terms

The foundational software layer that pools compute and storage across cluster nodes and presents them as shared resources. AI deployments size GPU hosts, storage tiering, and network fabric on top of this distributed platform.

Nutanix's native hypervisor that runs virtual machines directly on cluster nodes. GPU passthrough and vGPU profiles are configured on VMs running under AHV, making it central to AI workload placement.

The per-cluster management interface for configuration, health, and alerts. Operators use it for local GPU host settings, storage container policies, and cluster-level monitoring of AI infrastructure.

A multi-cluster management plane that aggregates Prism Elements and adds categories, policies, and centralized reporting. Multitenant AI environments rely on it for scoping, RBAC, and cross-cluster visibility.

The software-defined storage layer that spreads data across all nodes with local SSD and HDD tiers. It supplies the throughput and resilience that training data pipelines and checkpoints depend on.

A logical storage grouping on DSF with its own policies for compression, deduplication, and replication factor. AI datasets are usually placed in containers tuned for high sequential read and low dedup overhead.

A data protection scheme that stores parity fragments instead of full replicas, trading some write cost for lower capacity overhead. Large immutable AI datasets often benefit from it compared with replica-based protection.

The number of copies of each data extent that DSF maintains across nodes. Raising it improves durability for critical model artifacts but consumes capacity that could otherwise hold training data.

An S3-compatible object store that runs on Nutanix infrastructure and scales across clusters. It backs data lakes, model artifact repositories, and bucket-based access for AI pipelines.

An HTTP interface exposing bucket and object operations using the Amazon S3 API conventions. AI frameworks and tools point at this endpoint to read datasets and write checkpoints without code changes.

A scale-out file service providing SMB and NFS shares on the Nutanix platform. Shared NFS exports are common for multi-node training jobs that need a common dataset and checkpoint location.

A storage mount presented over the Network File System protocol and registered as a datastore for virtualization or containers. It gives many GPU workers concurrent access to the same training corpus.

A database lifecycle management service that provisions, clones, and patches database engines on Nutanix. AI applications that need relational or vector-capable databases use it for fast test and dev copies.

A space-efficient copy of a database created from a snapshot or gold image, often writable almost instantly. Teams use clones to refresh feature stores or evaluation datasets without duplicating storage.

A centralized repository holding raw and processed data at scale, usually on object storage. AI teams land training corpora here before curation, labeling, and feature engineering.

Nutanix's curated Kubernetes distribution for deploying and managing clusters across on-premises and cloud environments. It hosts containerized inference services, notebooks, and ML operators.

A Kubernetes operator that automates driver, container runtime, and device plugin installation on GPU nodes. It makes GPUs schedulable in NKP clusters without manual node configuration.

A Kubernetes extension that advertises a hardware resource such as a GPU to the scheduler. Pods request GPUs through resource limits, and the plugin handles allocation on the node.

The process of assigning pods or VMs to hosts with available GPU capacity based on requests, taints, and affinity rules. Correct configuration prevents contention and idle accelerators in shared clusters.

A scheduling rule that constrains which nodes a workload may run on, based on node labels. AI jobs use it to pin work to GPU models or to isolate tenants on dedicated hardware.

A paired mechanism where a node repels pods unless they explicitly tolerate its taint. Operators taint GPU nodes so only accelerator-consuming workloads land there, preserving scarce capacity.

A configuration that gives a virtual machine direct, exclusive control of a physical GPU. It delivers near-native performance for training but prevents sharing that card with other VMs.

A technology that partitions a physical GPU into virtual slices assigned to multiple virtual machines. It improves utilization for lighter inference or desktop workloads while isolating each consumer.

A hardware partitioning feature on certain NVIDIA GPUs that splits one card into isolated instances with dedicated memory and compute. It suits small inference tasks that do not need a full accelerator.

A reference architecture and deployment model for running generative AI models and supporting services on Nutanix infrastructure. It bundles validated compute, storage, and software guidance for self-hosted inference.

Running a trained model to produce predictions or generated output for requests. Inference workloads are latency-sensitive and often served from containers with GPU sharing enabled.

The runtime layer that loads a model, exposes an API endpoint, and handles batching and scaling of requests. Serving platforms run on Kubernetes and pull model artifacts from object storage.

A versioned catalog of trained model artifacts with metadata, lineage, and stage labels. It lets teams promote a validated model to production while keeping prior versions reproducible.

A validated configuration of compute, GPU, storage, and networking sized for a specific AI use case. It removes guesswork by documenting tested component combinations and capacity headroom.

A documented design pattern describing components, sizing, and interconnections for a workload class. It gives architects a starting baseline that they adapt to site-specific requirements.

Key-value metadata labels applied to VMs, clusters, or other entities for grouping. Policies and role assignments can target categories, enabling automated and tenant-scoped governance in AI environments.

An access model that grants permissions to roles and assigns roles to users or groups. Multitenant AI platforms scope roles to projects or categories so tenants see only their own resources.

Separating Kubernetes workloads into distinct namespaces with their own quotas, policies, and network rules. It prevents one team's training job from consuming another team's GPU or storage budget.

A Kubernetes limit on aggregate CPU, memory, GPU, or storage that a namespace may consume. Quotas keep a single AI team from monopolizing shared accelerator capacity.

Protecting stored data by encrypting volumes or buckets so physical access to media does not expose content. It is a baseline control for regulated training data and model artifacts.

A Kubernetes storage abstraction that outlives individual pods and binds to a backing datastore. Model caches and datasets mount as persistent volumes so restarted pods keep their data.

A set of nodes managed as a single unit that shares storage and compute resources. AI capacity planning starts at cluster granularity because GPU hosts and storage scale together.

The practice of tracking versions and applying upgrades, patches, and retirement to infrastructure and software components. AI environments need coordinated updates across hypervisor, Kubernetes, drivers, and model servers.

A Nutanix capability that inventories software components and performs automated updates with compatibility checks. It reduces risk when upgrading GPU drivers or platform firmware in production AI clusters.