SIAgentNews.com
← All news

NVIDIA open-sources NVKE and NVKR for GPU-managed Kubernetes

NVIDIA released two open-source projects to help organizations build and manage their own managed Kubernetes services on GPU hardware.

· First reported by marktechpost.com

1.36Kubernetes version for conformance testing

The short version

  • NVKE serves as the control plane for managed Kubernetes, while NVKR handles the building of OS and container images.
  • Both projects are released under the Apache 2.0 license and are currently in public preview.
  • The system supports both amd64 and arm64 architectures and integrates with NVIDIA Infra Controller (NICo).
  • Security features include SPIFFE identity attestation using TPMs and per-cluster etcd encryption.
  • Local evaluation of the software requires 6 CPUs and 12 GiB of RAM but does not require a GPU.

Why it matters

For organizations building private AI infrastructure, these tools provide a standardized way to manage GPU clusters without relying on public cloud providers. It moves the responsibility of 'managed' services onto the user's own hardware while providing the automation typically found in cloud environments.

A self-hosting AI factory

NVIDIA has introduced two new open-source projects designed to help users run managed Kubernetes services on their own GPU hardware. These projects, known as NVIDIA Kubernetes Engine (NVKE) and NVIDIA Kubernetes Recipes (NVKR), were recently announced as parts of the NVIDIA DSX AI Factory Platform.

NVKE acts as the control plane for the service, providing a way for operators to manage clusters via a REST API or a command-line tool. In contrast, NVKR functions as a build system that creates the necessary OS images, container images, and software packages required for these clusters to function.

Both projects are currently in public preview and are released under the Apache 2.0 license. While they are intended for GPU infrastructure, they are designed to be open and attested, providing a path for users to run upstream Kubernetes on their own hardware.

How the architecture works

The deployment process follows a specific five-stage workflow starting with a bootstrap phase. Users use a tool called nvkectl to start a temporary single-node cluster, which then requests three servers from the NVIDIA Infra Controller (NICo) to boot the management cluster.

Once the management cluster is active, the system handles machine provisioning through a Cluster API provider. This connects the hardware lifecycle to the NVKE software. Customers then pick specific release sets that define validated combinations of software versions and specific GPU generations.

Each customer cluster gets its own dedicated control plane, etcd, and node pools. The system includes several companion projects, such as AI Cluster Runtime (AICR) for GPU operations and NVSentinel for automated break-fix tasks. It also supports OpenTelemetry for exporting logs, metrics, and traces.

Security and hardware support

Security is a primary focus for the new stack, particularly regarding node identity. The system requires a node to present an attested SPIFFE identity before it can receive credentials to join the cluster. This process utilizes a Trusted Platform Module (TPM) to ensure hardware-level security.

The software is compatible with both amd64 and arm64 architectures. It is designed to run on bare-metal NVIDIA GPU servers provisioned through the NVIDIA Infra Controller. Additionally, per-cluster encryption for etcd is available to protect sensitive data within the management plane.

However, because the projects are in public preview, there are some current limitations. As marktechpost.com reported, the NVKE API serves plain HTTP on port 30080 with authentication turned off by default after installation. There is also currently no management-plane upgrade or backup procedure provided.

Market adoption and comparison

Several cloud partners and software companies are already beginning to integrate these projects into their offerings. For example, Firmus and SK Telecom are integrating the tools into their AI factories, while Mirantis and Rafay are adding components to their managed Kubernetes products.

When compared to other managed Kubernetes solutions like Gardener or Kamaji, NVKE stands out by providing a dedicated control plane for every customer. While Gardener runs as pods in seed clusters, NVKE gives each cluster its own dedicated infrastructure components for better isolation.

NVKR also provides a significant advantage in software provenance. It attaches signatures, Software Bill of Materials (SBOMs), and provenance to the OCI references it produces. This ensures that the software running on the GPU infrastructure is verified and traceable from the build stage forward.

Getting started with NVKE

Users interested in testing the software can perform a local evaluation without needing an NVIDIA account or a GPU. This quickstart requires a single Linux host or virtual machine with at least 6 CPUs, 12 GiB of free memory, and 100 GiB of disk space. It also requires Docker and virtual TPM support.

For a production-grade installation, the requirements are more extensive. A real deployment requires a NICo deployment with enrolled capacity, a build host equipped with KVM, an HTTP server, and a registry. Because NVKR does not have a qualified release yet, users should follow the official documentation for the latest updates.

The current status of the project is best described as a preview for developers. While it offers a robust foundation for GPU clouds, the lack of production-ready security and lifecycle features means it is currently best suited for evaluation and early-stage deployment rather than immediate production use.

What to watch

  • The release of a qualified NVKR release for production use.
  • Updates to the NVKE API to include default authentication and upgrade procedures.
  • More integrations from cloud partners like Mirantis and Rafay.

Try it

Run a local quickstart on a Linux host with 6 CPUs and 12 GiB RAM to test the software without needing a GPU.

Sources and links

First reported by marktechpost.com. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news