{"id":"storage","level":"datacenter","name":"Dataset and checkpoint storage hierarchy","shortName":"Storage","role":"Persists training data, model state, and results beyond accelerator working memory.","description":"AI storage provides the durable data behind repeated computation. Training reads samples and metadata, while checkpointing writes snapshots of model and optimizer state. These workloads stress different properties: many small operations, high sequential bandwidth, capacity, and recovery behavior may all matter. Local storage can stage data near a worker, while shared systems provide access across jobs and nodes. The atlas presents those layers separately from GPU HBM and CPU memory; storing a dataset does not mean every byte fits in accelerator memory at once. A well-supplied compute pipeline overlaps reading and preparation with execution. If storage cannot meet the workload's access pattern, expensive arithmetic hardware can wait idle.","specs":[{"label":"Durable contents","value":"Datasets, checkpoints, outputs","context":"Different access patterns may require different designs."},{"label":"Performance dimensions","value":"Bandwidth, IOPS, latency","context":"Capacity alone does not specify performance."},{"label":"Hierarchy","value":"Local staging + shared persistence","context":"A conceptual architecture, not one mandatory product."}],"interfaces":[{"name":"Data ingestion and checkpoints","direction":"in","signal":"Persistent records and model-state snapshots"},{"name":"Workers and recovery","direction":"out","signal":"Samples and restored state"}],"tradeoff":"Caching and staging improve locality. Capacity, consistency, and recovery requirements add complexity.","whyAI":"Training needs a sustained supply of examples and recoverable state. Storage performance influences useful accelerator utilization.","related":["checkpoint","leaf-switch","scheduler","serving","training-data"],"sources":[{"title":"NVIDIA — High-Performance Storage Architecture","url":"https://docs.nvidia.com/dgx-superpod/reference-architecture/scalable-infrastructure-b300/latest/storage-architecture.html"}],"scale":"Device-to-cluster storage hierarchy; capacities are workload-dependent","kind":"storage","mechanism":["Persist datasets and model-state files in suitable storage tiers.","Stage or stream the needed portions to active workers.","Write results and recoverable checkpoints back to durable storage."],"misconception":"Storage capacity is not the same as storage throughput, and durable storage is not GPU working memory.","check":{"question":"Why can a large storage system still starve GPUs?","options":["Its access latency or throughput may not match the workload","Large capacity guarantees unlimited I/O","GPUs cannot read data from any software pipeline"],"answer":0,"explanation":"Capacity describes how much fits. Sustained delivery depends on access patterns and the performance of the full data path."},"branch":"Data pipeline","subbranch":"Persistent I/O","science":[{"title":"Bandwidth and I/O operations","equation":"B ≈ IOPS × S_IO","explanation":"For a workload with average transfer size S_IO, operations per second and transfer size together determine an approximate byte rate. Small random requests and large sequential transfers stress different limits.","assumptions":"Average-rate relation over a defined workload; caching, metadata operations, concurrency, and latency alter observed performance."}],"history":[],"learningObjective":"Separate bytes read from storage from model working-set memory.","engineeringExample":"If 10 GB/s feeds a preprocessing stage for 60 seconds, 600 GB crosses that boundary; it need not reside in HBM simultaneously.","evidenceNotes":"IOPS and sequential bandwidth have different workload meanings.","visualFamily":"hall-system","prerequisites":["nvme-storage"],"labId":"library","scope":"Use the manufacturing investigation to trace compatible inputs, assembly and operating constraints; no current standalone calculator models this lesson’s full mechanism.","investigationId":"manufacturing","labScope":"Use the manufacturing investigation to trace compatible inputs, assembly and operating constraints; no current standalone calculator models this lesson’s full mechanism."}