Why is fast storage important for some training pipelines?

1. Input throughput must keep accelerators supplied with data
2. All training datasets fit in CPU registers
3. Fast storage replaces all network communication
4. Storage speed determines every model's statistical accuracy directly