A lighthouse-foghorn classifier has a few hundred rows. An engineer copies a multi-GPU distributed recipe from a larger job and watches communication overhead exceed compute. What should they do?
Select an answer to reveal the explanation.
Short Explanation
Think of a foghorn classifier with a few hundred rows and a copied multi-GPU recipe. Skip distributed training. Communication overhead will eat the compute. A few hundred rows do not pay back the sync.
Full Explanation
Skip distributed training when the dataset is too small for synchronization cost to pay back. Copying a multi-GPU recipe onto a few hundred rows wastes time on communication. Rekognition is not that decision, and forcing data parallel on every job is the same trap.