Skip to main content
Version: main 🚧

NVIDIA DSX OS with vCluster Platform

Turn NVIDIA infrastructure and software into services your tenants can use. vCluster Platform connects NICo-managed hardware, GPU health tooling, and NVIDIA application stacks to the same tenant, project, and cluster workflows. These integrations help you provision capacity, observe the fleet, and deliver managed Kubernetes, AI platforms, and inference services on your infrastructure.

NVIDIA DSX OS brings together modular software for operating AI infrastructure. Platform incorporates selected components as first-party provisioning integrations, bundled application templates, and Certified Stacks. The NVIDIA software runs alongside Platform's tenant isolation, access controls, and cluster management.

NVIDIA software, Platform workflowsThree independently configured Platform integrations: provision hardware with vMetal and NICo, observe GPUs with included NVIDIA health templates, and deliver applications with Run:ai and Dynamo Certified Stacks.NVIDIA software, Platform workflowsvCluster PlatformIncluded integrations, configured for your environmentProvisionvMetal + NVIDIA NICoMachines and nodesObserveNVIDIA GPU OperatorFleet Intelligence AgentNVSentinelDeliverCertified StacksNVIDIA Run:aiNVIDIA Dynamo
Provision hardware, observe GPU health, and deliver NVIDIA applications through Platform workflows. Configure each integration for your environment.

Start with the outcome you need​

As a platform admin, connect your hardware and configure GPU health monitoring. Then choose the services you want to offer your tenants.

Provision NVIDIA hardware​

NVIDIA Infra Controller (NICo) discovers and provisions bare metal and manages its infrastructure lifecycle. Deploy NICo separately, then connect it to the platform through the included vMetal NICo driver. The driver connects NICo-managed hardware to the platform's machine and node provisioning APIs. You keep NICo at the infrastructure layer and use Platform to control who can consume its capacity.

This path supports two service offerings:

  • Bare metal instances: allocate NICo-managed servers to a tenant or project as instances.
  • Cluster nodes: provision capacity for a cluster and use Platform's node workflows to bring it into service.

The platform admin configures the NICo connection, available node types, images, and network environments. Tenant and project policies then determine which capacity users can request. For tenants with Kubernetes access or arbitrary workload execution, use private nodes to give each cluster its own workers.

Start with vMetal, machine types and node types, and cluster attachment. Then connect NVIDIA Infra Controller and provision instances and cluster nodes with NICo.

For a wider view of the infrastructure workflows, read the vMetal deep dive. That walkthrough uses the Metal3 path. It explains the common vMetal model rather than validating the NICo driver.

Monitor GPU health​

GPU capacity needs a health signal as well as a provisioning API. Platform includes application templates for the NVIDIA GPU Operator, Fleet Intelligence Agent, and NVSentinel. Deploy them into the cluster that owns the GPU nodes, then configure Fleet Observability to collect and query their telemetry.

Included integrationWhat you getGuide
NVIDIA GPU Operator templateDriver, device plugin, DCGM, and GPU metric exporter configuration.GPU observability templates
Fleet Intelligence Agent templateNVIDIA node-level health and GPU metric collection.Deployment order and metric behavior
NVSentinel templateGPU fault detection, with configurable quarantine and remediation components.Observe GPU nodes with NVSentinel
Fleet Observability integrationA Platform-managed gateway and scoped query path to your configured telemetry backend.Fleet Observability quickstart

The templates are included. Deployment is explicit. Enable the Argo CD integration and select the templates for your target cluster. Choose one owner for GPU drivers so a provider-managed installation and NVIDIA GPU Operator don't compete for the same node. Configure the collectors, gateway, and backend before expecting fleet dashboards to show data.

This creates a path from a GPU node's health signal to an operator's view across clusters. Platform-owned labels identify the project and cluster, so operators can investigate fleet-wide issues without losing their origin. NVIDIA Fleet Intelligence Agent collects hardware signals. Platform Fleet Observability provides the configured collection and query path.

By default, the NVSentinel template only observes. Quarantine, drain, and remediation are opt-in. Use the NVSentinel guide to enable actions and check their dependencies. Automatic GPU-node replacement through vMetal remains outside this documented integration.

Fleet Observability walkthrough shows the collection, gateway, backend, and dashboard setup using an earlier Platform release. Use the linked docs for the current procedure and the blog post for a worked example.

Use first-party stacks included with Platform​

NVIDIA Run:ai and Dynamo integrations are included in the Platform v5.0 Certified Stack catalog. You select the integration, supply its environment-specific parameters, and install it through Platform's Stack workflows. You don't need to author the application dependency graph from scratch.

vCluster maintains the orchestration templates and bundles them with Platform. NVIDIA develops the underlying software. First-party integration means the Platform installation and lifecycle use native Apps, StackTemplates, and StackInstances. It doesn't change NVIDIA's licensing, registry access, or hardware requirements.

Included Certified StackService you can buildIntegration path
NVIDIA Run:aiGPU scheduling and workload management for AI teams.Dedicated and central deployment models
NVIDIA DynamoA distributed model-serving runtime on a GPU cluster.Install the Dynamo runtime and see a worked deployment

A Stack adds the coordination that a collection of independent chart installs leaves to the operator:

  • Ordered deployment: tasks start after their dependencies are healthy.
  • Shared configuration: parameters and task outputs connect the applications that belong together.
  • Visible progress: aggregate health and task-level messages show where an installation needs attention.
  • Repeatable delivery: include a Stack in a cluster template or install it into an existing destination.
  • Managed changes: update a StackInstance's parameters and reconcile the application's desired configuration.

The catalog is pinned to each Platform release. Bundled Certified StackTemplates are read-only in the platform UI. Keep the bundled resource when using its maintained integration, or create a separate custom template when you need to change it. For setup, permissions, and lifecycle, see Certified Stacks and Use a Stack.

Bundled doesn't mean automatically installed into every cluster. You still supply GPU capacity, storage, registry credentials, and workload-specific configuration. For Dynamo, complete an inference request before treating a healthy installation as a working model-serving service.

Dynamo deep dive walks through an inference deployment on private GPU nodes. The blog post demonstrates the application. The bundled Stack provides the Platform delivery path.

Validate GPU configurations​

The DSX ecosystem also provides tools for checking what you are about to hand to a tenant. These engineering walkthroughs show how to check GPU software configuration and test workloads on provisioned nodes:

These are worked evaluations, distinct from the first-party catalog entries above. A readiness report provides evidence about the tested nodes and workload. It doesn't automatically admit a cluster to service, quarantine a node, or initiate replacement through Platform. The blog posts use earlier software versions. Repeat the checks on your target versions and hardware.

Put the integrations together​

Together, the integrations take you from raw capacity to a tenant service:

  1. Make hardware consumable. Configure a NICo site through the vMetal driver and expose the capacity through node types and tenant policies.
  2. Prepare an isolated cluster. Give the tenant a cluster with private GPU nodes and configure its GPU software owner.
  3. Establish health visibility. Deploy the GPU templates and complete the Fleet Observability collection and query path.
  4. Deliver the service. Install a Run:ai or Dynamo Stack for the scheduling or serving environment the tenant needs.
  5. Confirm the outcome. Inspect the assigned nodes, health signals, Stack task results, and an actual tenant workload.

Each step follows its own guide. The sequence isn't an automatic end-to-end remediation loop.

Further integrations​

NVIDIA Kubernetes Recipes (NVKR) builds and publishes artifacts from pinned inputs. NVIDIA Kubernetes Engine (NVKE) provides its own management plane for provisioning Kubernetes clusters. For teams evaluating those projects, Platform offers two additional integration paths:

The installation Stack verifies the chosen release before deploying management services. It doesn't install NICo, Cluster API, or the downstream provisioning configuration. Connecting an NVKE-created cluster retains NVKE's ownership of its provisioning, upgrades, scaling, and deletion.

The NVKE management-plane Stack is included in the Platform v5.0 Certified Stack catalog. Its guide also provides a pinned source fallback for builds without the template. NVKE and NVKR are upstream previews. Signature verification authenticates artifacts. It doesn't establish production readiness.