CDI: The Future of Specialized Hardware in Containers
summary
CDI defines a declarative, CNCF‑sponsored spec that maps complex hardware access to OCI runtime spec edits (device nodes, mounts, env and hooks), letting vendors ship a YAML/JSON device description instead of runtime‑specific plugins. Demo uses NVIDIA ctk to generate specs on Docker 26, shows an A100 server with 8 GPUs → 17 CDI devices (indices, UUIDs, an “all” meta device), and runs a PyTorch MNIST job on GPU0 successfully. Notes: runtimes support varies (Podman/cri-o enabled, Docker/containerd opt‑in), and spec staleness across driver upgrades remains an operational concern.
AI-written summary · about
chapters (14)
- 0:00Intro
- 1:29ThirdParty Device Support
- 3:12Consequences
- 4:27NVIDIA GPU Example
- 6:05CDI
- 8:39Nvidia GPU
- 10:15Intel GPU
- 10:53Device Injection Flow
- 14:03Benefits
- 15:50Demo
- 21:30Where to use CDI
- 23:45Whats next
- 25:49How to get involved
- 26:41Questions
related talks
05mTowards a Sustainable World Realized by Kubernetes and Composable...
36mcontainerd Deep Dive
36mExtending containerd
29mContainerd, What Does it Mean for Me - Scott Coulton, Puppet
43mContainerization in Mesos, Embracing the Standards
34mContainerd: Project Update and Deep Dive
05mLightning Talk: Managing Drivers in a Kubernetes Cluster
25mReverse Engineering Cloud Native: Interoperability and Community