Ignara AI/Docs/Storage Fabric
Architecture & ResearchUpdated June 2026

Storage Fabric

On this page

AI storage requirementsArchitectureCaching tiersTraining pipeline

AI storage requirements

AI training workloads have storage requirements that differ significantly from transactional or analytical workloads. Key characteristics include: high sequential read throughput, large object sizes (model checkpoints, dataset shards), and sensitivity to read latency (GPU starvation from slow data loading).

Architecture

The Fabric storage layer is organized as a distributed object store with a tiered caching architecture. Data is stored in a durable remote tier and cached locally on compute nodes using available NVMe and DRAM capacity.

Caching tiers

  • Remote storage — durable, high-capacity object storage for dataset and checkpoint persistence
  • NVMe cache — fast local SSD cache for frequently accessed data shards
  • DRAM cache — in-memory cache for hot data paths and streaming prefetch buffers

Training pipeline

The storage layer includes a training-aware pipeline that prefetches data shards in the order they will be consumed by the training job. This eliminates GPU idle time from data loading and keeps compute utilization high throughout the training run.

PreviousSchedulingNext Network Fabric