What is a common storage need for training at scale?
Why D is the answer
High sustained read throughput
GPUs must not wait for data.
In detail
- GPUs must not wait for data.
- GPUs must not wait for data.
Use case
An illustrative scenario, not a real incident.
Training at scale reads huge datasets, so storage needs high sustained throughput.
See it in different ways
flowchart TD A["AI storage"] --> B["High read throughput for training"] A --> C["Fast checkpoint writes"] A --> D["Many small files for metadata"]
mindmap
root((AI storage))
Reads
High throughput
Parallel file system
Writes
Fast checkpoints
Direct path
GPUDirect Storageflowchart LR S0["Storage"] --> S1["High read throughput"] --> S2["GPUs stay busy"]
Key point: Training needs sustained read throughput.
Why the other options are wrong
- A. Very slow, cheap archive access
- Archive access is too slow.
- B. A single local SATA disk
- One disk cannot feed a cluster.
- C. One small shared folder
- One small folder cannot feed a cluster.