Nvidia announced two open-source projects on 10 November for organisations that want to offer managed Kubernetes clusters on their own GPU infrastructure. Nvidia Kubernetes Engine (NVKE) provisions and operates the clusters, while Nvidia Kubernetes Recipes (NVKR) builds the software they run. Both are public previews on GitHub under the Apache 2.0 licence and form part of Nvidia’s DSX AI Factory Platform, which the company says helps organisations build and operate GPU infrastructure.
Key points
- NVKE provisions separate Kubernetes control planes, data stores and node pools for customer clusters.
- NVKR builds node images and other components from pinned software versions.
- Nvidia says preview APIs and behaviour may change and does not recommend the projects for production workloads.
NVKE assigns each customer a separate cluster
The announcement follows earlier accounts of the preview. LavX reported NVKE’s release on 6 November. Heise reported its unveiling on 7 November. Nvidia’s own announcement is dated 10 November and covers both NVKE and NVKR.
NVKE runs a management plane on a Kubernetes cluster that provisions customer workload clusters on bare-metal servers. Nvidia says each customer cluster receives its own control plane, its own etcd store for Kubernetes state and its own pools of machines. A temporary cluster on a bootstrap host starts the management cluster, then is removed once the management cluster’s nodes have joined. The management plane subsequently accepts cluster requests through an API or the nvkectl command-line tool.
The separation matters because Kubernetes needs somewhere to keep track of what should be running and which machines are available. Giving each customer cluster its own control plane and etcd store is rather like giving each building its own switchboard rather than routing every call through one shared desk. Nvidia says the arrangement is intended for operators offering customers managed clusters on GPU hardware they control.
Writing a work report with AI could run on a GPU cluster separate from clusters assigned to other customers, if the AI service used this infrastructure. That separation would concern the computers processing the request, rather than the way the report is written.
For organisations building such a service, NVKE provides the cluster-management software while leaving them in charge of their operations and customer experience, according to Nvidia. The projects sit within the same DSX platform in which Nvidia made the Run:ai platform open source, as AI Affairs previously reported.
NVKR fixes the inputs to node images
NVKR handles a different part of the job: preparing the software that NVKE installs. Nvidia says its recipes specify software versions and GPU generations, then build node images, container images, binaries and Helm charts while recording the inputs. An operator can therefore identify what went into an image and rebuild it from the recorded sources. The management cluster and different node pools can use different images.
In Nvidia’s walkthrough, NVKR builds a disk image that includes Kubernetes packages and a node agent. NVKE then uses that image to start servers supplied through Nvidia Infra Controller, or NICo, which manages the bare-metal machines. A proposed cluster specifies a Kubernetes version and separate configurations for its control-plane and GPU worker nodes. Nvidia says NVKE can use software built without NVKR, and NVKR can produce artifacts for other cluster-management systems.
Before a machine receives credentials to join a cluster, Nvidia says it must present an attested SPIFFE identity. SPIFFE provides an identity for software and machines, while SPIRE issues and verifies those identities. The walkthrough also describes a telemetry pipeline that collects logs, metrics and traces for export to a monitoring system. Nvidia says NVKE has been certified as a Kubernetes AI Platform and Kubernetes Installer after conformance testing on Kubernetes 1.36.
NVKE preview leaves management work with operators
Nvidia says NVKE and NVKR remain in public preview, with APIs, configuration and behaviour liable to change between releases. It does not recommend either project for production workloads. The distinction matters for an organisation considering a managed service: the walkthrough describes how to build an image and create a customer cluster, while Nvidia’s stated production recommendation remains separate from that procedure.
The management cluster does not upgrade, scale or replace its own nodes, and the preview has no procedure for backing up and restoring the management plane, heise reported. Its account says every authorised caller can create and delete all workload clusters because access controls for individual clusters are still missing.
For a local evaluation, heise described a quickstart that substitutes a mock controller for NICo and runs workload-cluster nodes as Docker containers. The Linux machine must have six or more CPUs, 12 GB of available RAM, 100 GB of available storage, Docker and the ability to run a virtualised TPM.
