Status: Proposal — not implemented. Containers only in v1; microVM passthrough explicitly deferred.
Adds registered devices: the operator explicitly registers a host device (PCI device, /dev/dri/* render node, USB-serial — anything with a /dev node) under a logical name, and workloads reference it by name in their spec. The container driver translates that into OCI runtime bind + cgroup device-allow. Manifests stay portable (name, not host PCIe address). No host-side AI stack — the container brings its own oneAPI/CUDA/etc. Sharing by default. Motivating workload: Intel Arc Pro B50 for inference.
📄 Full proposal: https://github.com/Project-Capsule/capsule/blob/main/docs/pci-devices.md
Status: Proposal — not implemented. Containers only in v1; microVM passthrough explicitly deferred.
Adds registered devices: the operator explicitly registers a host device (PCI device,
/dev/dri/*render node, USB-serial — anything with a/devnode) under a logical name, and workloads reference it by name in their spec. The container driver translates that into OCI runtime bind + cgroup device-allow. Manifests stay portable (name, not host PCIe address). No host-side AI stack — the container brings its own oneAPI/CUDA/etc. Sharing by default. Motivating workload: Intel Arc Pro B50 for inference.📄 Full proposal: https://github.com/Project-Capsule/capsule/blob/main/docs/pci-devices.md