Diagnose the bottleneck first
A slow training job may be waiting on storage, preprocessing, or communication rather than computation. Profile a small representative run before adding accelerators. Compare input throughput with accelerator utilization. More workers can increase overhead if the dataset or model does not benefit from parallelism.
A training job has low GPU utilization and high input wait time. List two changes to investigate before buying a larger GPU.
Tune with a budget and a baseline
Choose a validation metric aligned with the business objective and define a search budget. Start with influential parameters and reasonable ranges. Use early stopping where appropriate, and keep the final test set out of tuning decisions. Record the improvement relative to the baseline and the additional training and serving costs.
Design a small tuning experiment for a tree-based classifier. Explain which validation metric you would maximize for a rare positive class.
Understand distributed training tradeoffs
Data parallelism distributes batches across model replicas. Model parallelism splits model computation across devices. Neither removes the need to consider communication overhead, memory limits, and checkpoint recovery. For foundation models, compare prompting, retrieval, and tuning against a fixed evaluation set before committing to a training strategy.
Describe a case where the model cannot fit on one device and a separate case where the model fits but training takes too long.
Your accelerator spends most of its time waiting for batches. What is the best first investigation?
Keep exploring
Use the official certification page and its linked resources to validate your coverage. For guided labs, explore the Google Skills learning path. Revisit the practical exercise and explain your decision in your own words before moving on.