A model's input preparation consumes most elapsed time while the GPU is idle. What should be addressed first?

1. Buy more identical GPUs without measuring the pipeline
2. The data preparation and feeding bottleneck
3. Disable all host-side metrics
4. Increase the model's checkpoint frequency only