The four layers of AI storage
Datasets, volumes, images, and shared training state — each on the right medium, all in the same data centers as your GPUs.
AI Object Storage
S3-compatible object store (EOS) for training datasets, model artifacts, and backups — versioning, lifecycle rules, replication, object lock, and pre-signed URLs, all driven by the standard S3 API and toolchain. Mounts into Kubernetes via open-source Datashim and feeds TIR datasets directly.
Explore Object Storage →Block Storage
Attachable, expandable volumes for nodes — boot disks, databases, scratch space — with per-volume snapshots and restore. Kubernetes-native via the Ceph RBD CSI StorageClass, with LUKS encryption at rest.
Explore Block Storage →Container Registry
Private registry at registry.e2enetworks.net — standard Docker push/pull, token-based access control, vulnerability scanning, and replication. Wire it to Kubernetes with imagePullSecrets or serve TIR custom containers.
Explore Container Registry →Parallel File Storage
Shared high-throughput filesystem for multi-node GPU training — mounts across TIR instances and Slurm clusters for checkpoints and shared datasets, with read-only mount options for serving.
Explore PFS →Which storage for which job?
A quick decision guide for AI workloads.
| You need to… | Use | Why |
|---|---|---|
| Store training datasets, model weights, backups | AI Object Storage | S3 API, versioning, lifecycle, cheapest per GB |
| Give a VM or database fast persistent disk | Block Storage | Attach/expand volumes, snapshots, CSI for K8s |
| Ship and pull private container images | Container Registry | Docker-standard workflow, scanning, tokens |
| Feed a multi-node training run or share checkpoints | Parallel File Storage | Shared POSIX-style mount across GPU nodes |
| Serve a model behind an API | TIR Endpoints + Object Storage | Pull weights from EOS, autoscale workers |
Standards-based means portable
Every storage product speaks a protocol you already use — so moving in is easy, and so is moving out.
S3 API
Object storage works with the standard S3 toolchain — access key, secret, endpoint. No proprietary SDK.
Docker / OCI toolchain
Standard docker push/pull against the registry; sha256-digest image refs; Kubernetes imagePullSecrets.
Kubernetes CSI
Block volumes provision dynamically via the csi-rbd-sc StorageClass — plain PVC manifests, no custom controllers.
Filesystem mounts
PFS and shared FS mount as ordinary paths on TIR instances and clusters — your training code doesn't change.
Full API surface documented with OpenAPI specs at docs.e2enetworks.com/api.
Put your data next to your GPUs
Same data centers, same INR invoice, standard protocols.