Rank #06 · Daily · Infra
NVKE: Management Plane for GPU Kubernetes on Bare Metal (Preview)
Install an NVKE management plane on your NVIDIA GPU iron; create isolated Kubernetes workload clusters for tenants through the NVKE API / nvkectl.
What it does
NVIDIA Kubernetes Engine (NVKE) (repo under dsx-ai-factory/nvke, not the NVIDIA GitHub org) is a management plane for running a managed-Kubernetes-like service on your own NVIDIA GPU infrastructure. GitHub description: "NVIDIA Kubernetes Engine (NVKE). A management plane for GPU Kubernetes on bare metal." Docs: docs.nvidia.com/nvke (also in-repo docs/).
From the README (preview):
- Not a Kubernetes distro and not "just another Cluster API flavour." Framing: a self-hosting factory — seed → management cluster → many isolated workload clusters, each with its own control plane, etcd, and node pools.
- Creates/scales/deletes node pools, upgrades Kubernetes, replaces nodes (with preview limits documented).
- Node identity via SPIFFE/SPIRE + TPM attestation before join credentials; optional etcd encryption per cluster.
- Infrastructure provider at preview: NVIDIA Infra Controller (NICo) only. Companion NVKR builds/signs node images and components.
- Operator surfaces: NVKE API /
nvkectl; build produces binaries (agent,controller,nvkectl,management-engine,tenant-api,operator-api, …) with Go 1.27.3 pinned. - Local quickstart can evaluate on simulated machines without GPU hardware (Linux amd64/arm64; macOS via Colima guide) — Docker + vTPM + listed tools; ~6 CPUs / 12 GiB RAM / 100 GiB disk suggested.
Preview status (important). README: APIs/config/behavior can change; not recommended for production. After install, API auth is off and uses plain HTTP until you configure protected access. Management cluster does not self-upgrade/scale/replace; no management-plane backup/restore procedure yet. FDE node images are evaluation-only. Read docs/overview/preview-status.md before choosing a path.
Who it's for
- Platform teams evaluating on-prem / bare-metal GPU Kubernetes with a multi-tenant management plane on NVIDIA stacks.
- Contributors and researchers exploring NVKE/NICo/NVKR integration before production readiness.
- Operators who can accept preview limitations and supply machines, networking, images, and NICo access for real installs.
Not a fit if you need a turnkey production GPU PaaS tomorrow, or only want a single kubeadm cluster without a management plane.
Stars / language / license
859 stars · Go · Apache-2.0
Data verified: 2026-11-09 09:16 Asia/Shanghai (GitHub API get_repository + README)
Trending note: +345 stars today (captured 09:10 Asia/Shanghai)
Category tags
Infra · Kubernetes · GPU · bare-metal · NVIDIA · Cluster-API · Go · management-plane
How it differs from similar projects
Compared with Rancher / OpenShift / Tanzu-class management planes. Similar "management cluster provisions workload clusters" shape, but NVKE is NVIDIA GPU bare-metal / NICo-specific and explicitly preview.
Compared with vanilla Cluster API + GPU operator glue. CAPI is a building block. NVKE packages a product workflow (seed, management plane, attestation, nvkectl, observability hooks) aimed at GPU fleet operators — still early.
Compared with cloud GKE/EKS/AKS GPU node pools. Those are hosted. NVKE targets your machines and registries.
Caveats. Young star count relative to peers on this list; treat capability claims as README-scoped. Confirm NICo availability and supported platform matrix before investing. No TidyHow deployment verification.