Key takeaways

  • Run a managed Kubernetes service on your own NVIDIA GPU machines, with a separate cluster for each customer.
  • Every workload cluster gets its own control plane, etcd and node pools, so customers stay isolated from each other.
  • Nodes must prove their identity with a TPM before they receive join credentials.
  • You can evaluate it on simulated machines, with no GPU hardware and no NVIDIA account.
  • The catches: it is a preview, API authentication is off after install, and only one infrastructure provider works.

What it is

NVIDIA Kubernetes Engine, or NVKE, is a management plane for GPU Kubernetes on bare metal, released under the Apache-2.0 licence. The code is on GitHub at dsx-ai-factory/nvke.

The problem it solves

Platform teams that own GPU servers often want to hand out Kubernetes clusters to internal customers, rather than share one cluster between everyone. NVKE is meant for that job: you operate the management plane, and your customers use their clusters through the Kubernetes API and kubectl.

The README calls it a self-hosting factory for Kubernetes clusters. It is not a Kubernetes distribution and not a Cluster API flavour.

What you get

  • Isolation: each workload cluster has its own Kubernetes control plane, etcd and node pools.
  • Lifecycle tools: node pools can be created, scaled and deleted, and Kubernetes can be upgraded, with preview limits on replacement and upgrades.
  • Node identity: nodes attest through SPIFFE/SPIRE using their TPM before receiving join credentials, and etcd encryption is available per cluster.
  • A build link: the companion project NVKR builds node images and NVKE components from pinned inputs.
  • A way to look first: a local quickstart runs a development management plane and a workload cluster on simulated machines, with no NVIDIA account.

The catches

  • Maturity: NVKE is available as a preview, and the project says it does not recommend it for production workloads.
  • Security defaults: API authentication is off and the API uses plain HTTP after installation, so operators must add protection before admitting callers.
  • Operations gaps: there is no management-plane upgrade or backup and restore procedure.
  • Platform limits: NICo is the sole compatible infrastructure provider at preview, and the local quickstart needs a Linux host or VM.
  • Evaluation only: FDE node images are for evaluation and provide no data-at-rest confidentiality.

What we did not test

We did not run NVKE, build it, create a cluster or use the local quickstart. We also did not check the NICo integration, the node attestation or any claim about isolation. Every claim here comes from a literal quote in the project's README or the GitHub API, and the feature descriptions are the project's own.

Verdict

Editors' view: this is for platform and infrastructure teams who run their own NVIDIA GPU machines and want to see how a managed Kubernetes service could be built on them. Everyone else should treat it as a design to read about. Given the preview status and the security defaults, keep any trial on simulated machines.

Repository
dsx-ai-factory/nvke
Go · Apache-2.0
View on GitHub ↗

Sources

Need help adopting this, or have a software project in mind?

Tell us what you are building. The initial consultation is free. Email [email protected] or follow @zolnaltd.

Start a conversation →