NVIDIA Kubernetes Engine (NVKE) Infrastructure Management
A managed Kubernetes service platform for operating GPU-accelerated infrastructure clusters.
1Platform Management Plane
NVIDIA Kubernetes Engine (NVKE) serves as a management plane designed for platform teams to deploy and operate Kubernetes clusters on private GPU hardware. It enables the creation of isolated workload clusters, each featuring its own control plane, etcd storage, and node pools, allowing customers to interact via standard Kubernetes APIs.
2Infrastructure Provider Integration
The system currently relies on the NVIDIA Infra Controller (NICo) as its sole compatible infrastructure provider for provisioning machines. Operators are responsible for supplying the underlying hardware, networking, node images, and registries. This architecture requires specific configuration of access controls before admitting external callers to the management plane environment.
3Security and Identity
NVKE establishes node identity through the SPIFFE/SPIRE framework, utilizing Trusted Platform Module (TPM) hardware for attestation before granting join credentials. While the system supports per-cluster etcd data encryption, the current preview version lacks production-grade data-at-rest confidentiality features, making it suitable primarily for evaluation rather than sensitive production workloads.
4Lifecycle Management Capabilities
The platform automates the full lifecycle of workload clusters, including creation, scaling of node pools, and Kubernetes version upgrades. Users can manage these clusters through the NVKE API or the nvkectl command-line tool. However, the management plane itself does not currently support self-upgrading, automated scaling, or native backup and restore procedures.
5Development and Build Process
The project utilizes NVIDIA Kubernetes Recipes (NVKR) to build node images and system components from pinned inputs, ensuring consistent release artifacts. Developers can build the workspace from source using Go 1.27.3 and GNU Make 4. The build process generates nine distinct binaries, including the agent, controller, and tenant-api modules.
6Evaluation and Deployment
A local quickstart option allows users to evaluate NVKE on simulated machines without requiring dedicated GPU hardware. This setup runs on Linux hosts or macOS via Colima, requiring at least 6 CPUs, 12 GiB of memory, and 100 GiB of disk space to successfully confirm that cluster nodes report as ready.
7Architectural Design Philosophy
NVKE is explicitly defined as a self-hosting factory for Kubernetes clusters rather than a traditional distribution or Cluster API flavor. It employs a seed cluster to bootstrap the management plane, which subsequently provisions and manages multiple isolated workload clusters on demand, providing a structured approach to multi-tenant GPU infrastructure operations.
8Current Preview Limitations
As a preview release, NVKE is subject to significant changes in APIs, configuration, and behavioral patterns between versions. The system currently operates with API authentication disabled and uses plain HTTP, necessitating manual security hardening. Organizations should review all known limitations and documentation before integrating this tool into their infrastructure workflows.