Architecture decision
Profile the slowest stage before adding GPU capacity. A data-loading bottleneck can leave an accelerator idle.
GLOBALATLAS AI · Machine learning infrastructure
Machine learning infrastructure
Separate data preparation, training and serving; these stages need different compute and storage profiles.
Data preparation, training and evaluation
- 01
CPU pipeline for validation and preprocessing
- 02
Training host and persistent checkpoints
- 03
Evaluation dataset isolated from training inputs
Configuration requirements
Use CPU for preparation, GPU when the framework supports acceleration, and persistent storage for datasets and checkpoints. Measure the training batch before choosing VRAM.