26 lines
986 B
Markdown
26 lines
986 B
Markdown
|
---
|
||
|
|
title: NVIDIA Device Plugin
|
||
|
|
modified: 2026-03-27
|
||
|
|
tags:
|
||
|
|
- service
|
||
|
|
- gpu
|
||
|
|
---
|
||
|
|
|
||
|
|
# NVIDIA Device Plugin
|
||
|
|
|
||
|
|
Kubernetes device plugin that exposes NVIDIA GPUs to pods on [[ringtail]]. Required for GPU workloads like [[frigate]] (object detection) and [[ollama]] (LLM inference).
|
||
|
|
|
||
|
|
## Quick Reference
|
||
|
|
|
||
|
|
| Property | Value |
|
||
|
|
|----------|-------|
|
||
|
|
| **Namespace** | `nvidia-device-plugin` |
|
||
|
|
| **Image** | `nvcr.io/nvidia/k8s-device-plugin` |
|
||
|
|
| **Upstream** | https://github.com/NVIDIA/k8s-device-plugin |
|
||
|
|
| **Manifests** | [argocd/manifests/nvidia-device-plugin/](https://forge.eblu.me/eblume/blumeops/src/branch/main/argocd/manifests/nvidia-device-plugin) |
|
||
|
|
|
||
|
|
## Architecture
|
||
|
|
|
||
|
|
Runs as a DaemonSet with `privileged` security context, mounting the host's device-plugins socket, CDI specs, and NVIDIA driver libraries. A `RuntimeClass` named `nvidia` is defined for pods that need GPU access.
|
||
|
|
|
||
|
|
Time-slicing is configured with 2 replicas per GPU, allowing two pods to share a single physical GPU.
|