EXPLORINEER
en

DE — heise online

NVKE: Nvidia Launches Its Own Kubernetes Engine as Preview

Software2 min read

Nvidia has released NVKE, a new Kubernetes Engine, as a preview on GitHub. It lets companies that run their own GPU data centers offer Kubernetes to their customers as a managed service.

What happened

Nvidia showed its new Kubernetes Engine, called NVKE, as a preview on GitHub. Operators of GPU data centers can use it to give their customers Kubernetes as a managed service, similar to AWS EKS. A central management cluster creates separate Kubernetes clusters on bare-metal servers when customers ask for them. The code is free under the Apache 2.0 license, but Nvidia does not yet recommend NVKE for live production use.

Key facts

  • Each customer cluster gets its own control plane, its own etcd database, and its own node pools.
  • The management cluster needs an odd number of servers with TPM 2.0 chips, such as one, three, or five.
  • The NVKE API listens on unencrypted HTTP port 30080 on every management node.
  • New clusters use ovn-kubernetes for networking. Calico is only kept for existing installations.
  • NICo, the server-management layer, needs a Kubernetes cluster with at least three nodes.

How it works

NVKE sets up the servers without GPU and fabric software. Nvidia's GPU Operator then installs the drivers and related tools, and Nvidia's AI Cluster Runtime provides the setup recipes. Operators control everything with the nvkectl command-line tool or the NVKE REST API. They can also use the API to scale node pools and install Kubernetes upgrades. For high availability and upgrades, NVKE needs a virtual IP address, which it sets up with kube-vip.

Why it matters

NVKE is a sign that Nvidia wants to make its full hardware and software stack the default for running AI servers. But the preview still has gaps. It has no security roles, so any allowed caller can create or delete any cluster. It also has no tested way to upgrade to a newer NVKE release or to back up and restore the management cluster. Companies that want to try it should add their own TLS encryption and login checks first.

What to watch

  • Whether Nvidia adds role-based access control to the API.
  • Whether a safe upgrade and backup path arrives before NVKE is ready for production.
  • How NVKE compares with open-source tools such as Gardener and Kamaji, which run customer control planes as pods.

Glossary

Kubernetes
Software that runs and manages many containers across a group of computers.
NVKE
Nvidia's new Kubernetes Engine, which builds and manages customer Kubernetes clusters on GPU servers.
NICo
Nvidia Infra Controller, which manages bare-metal servers through software APIs instead of manual steps.
GPU Operator
Nvidia tool that installs the drivers and software a GPU needs inside a Kubernetes cluster.
TPM 2.0
Trusted Platform Module, a security chip on a server that stores keys and proves the machine's identity.

Same story, different eyes

Ad

Add 150 words. Learn 10 a day.

Dicter word books send only the words you learn to your review notes.

More in this topic

Softwareheise online

heise offer: From Skill to Plugin: Connecting AI with Tools and Data

A webinar titled "From Skill to Plugin: Connecting AI with Tools and Data" will be held on November 11 by heise KI PRO. It is aimed at professionals and managers already using generative AI for recurring tasks and covers how individual skills can be extended into plugins using connectors, including the MCP connection standard, read-only access versus actions that change content, and building and testing plugins with examples from ChatGPT, Claude, and Microsoft 365 Copilot. Tickets and further information are available on the heise conferences website.

SoftwareThe Register

Windows takes a break at UK motorway services

A roadside information display at Cherwell Valley motorway services in the UK stopped working and showed a Windows Automatic Repair screen after a restart problem. The Register reader who spotted it joked that the sign had taken the advice to take a break to heart. The article notes that the screen may point to a driver problem, a faulty update, or a hardware failure, and that the computer behind it will likely need to be accessed to fix it.

Softwareheise online

Rust across the board: Discord drops React and Electron

Discord plans to replace Electron and React in its desktop clients, with the core rewritten in Rust and the GPUI framework used for the interface. The desktop and mobile apps will share the same Rust-based application logic, and React Native will be discontinued. Discord says the new Rust base should use less than 300 MB of RAM, a figure not yet independently tested.

SoftwareArs Technica

Could potential AI clones of GTA VI force a quicker official port?

A former Rockstar tech director, Obbe Vermeij, says the threat of AI-assisted decompilation ports of GTA VI could push Rockstar to speed up the official PC version. He recalled that Rockstar has long prioritized consoles over PC, with PC ports of major games often arriving many months or years after console releases, despite PC versions once being favored by artists.