Loading
Loading
Object, file and block storage engineered for the I/O demands of AI training, inference and data-intensive enterprise applications — with tiered performance, encryption everywhere and policy-driven lifecycle management.
AI workloads place extreme demands on storage systems. Training a large language model requires sustained multi-GB/s write throughput for checkpointing, while inference serving demands microsecond-latency reads for KV-cache access. Dataset preparation involves scanning terabytes of unstructured data, and model repositories must serve artefacts to hundreds of nodes simultaneously. NOVACORE storage architecture addresses these patterns with a tiered approach: hot NVMe-based parallel filesystems for active training workloads, warm object storage for datasets and artefacts, cold archival tiers for compliance and backup, and ephemeral local NVMe scratch on every GPU node. Each tier is sized against the workload profile — not provisioned from a generic one-size-fits-all pool — and all tiers share a common encryption, access control and lifecycle management framework.
Scalable, S3-compatible object storage for datasets, model artefacts, training corpora and media at petabyte scale. Immutable object support for compliance, versioning for dataset lineage and lifecycle policies that automatically transition data between performance tiers based on access patterns.
High-performance parallel filesystems — Lustre, WEKA or GPFS — shared across training clusters for coordinated access to datasets, checkpoints and logs. Designed for the sustained multi-GB/s throughput that distributed training demands, with POSIX semantics and concurrent read/write from hundreds of GPU nodes.
Low-latency, high-IOPS volumes for databases, metadata stores, container registries and system disks. Configurable performance tiers with guaranteed IOPS and throughput. Snapshots, clones and cross-availability-zone replication for data protection and disaster recovery.
Policy-driven, tested backups with defined retention periods. Automated snapshot scheduling, incremental backup with deduplication, isolated backup networks and regular restore testing. Backup policies are defined per workload — production inference endpoints may require continuous data protection, while development datasets may use daily snapshots.
| Tier | Performance | Use case |
|---|---|---|
| Hot (NVMe parallel FS) | 100+ GB/s read, 50+ GB/s write | Active training, checkpointing |
| Warm (object storage) | Multi-GB/s, S3 API | Datasets, artefacts, logs |
| Cold (archive) | Cost-optimised, slower retrieval | Compliance, long-term retention |
| Local NVMe (ephemeral) | Per-node, 10–30 TB | Scratch, temporary data |
Secure AI and high-performance computing for enterprises, governments and research.