A public health department is training a large chest X-ray screening model, and the image dataset keeps growing each month as new scans are collected. Single-instance training runs are taking longer every cycle and are starting to delay releases. Which change best keeps training time manageable as the dataset scales?

Distributed training splits the workload across multiple compute instances or devices, most commonly through data parallelism, where each worker processes a different shard of the batch and gradients are synchronized, which lets total training throughput scale roughly with the number of workers added as the dataset grows. Increasing epochs on a single instance makes the problem worse, not better, because it multiplies the exact serial bottleneck that is already causing delays; more passes over the same hardware means more time, not less. Reducing resolution alone can shrink per-image compute cost, but it does not address the structural issue of a single instance being asked to process an ever-growing dataset, and it trades away image detail that a screening model may depend on. Moving storage from S3 to EFS changes where bytes live, not how compute is parallelized; storage choice can affect I/O throughput but does not by itself distribute the training computation across more workers. Scope note: distributed training introduces its own overhead from gradient synchronization, so scaling is sub-linear past a point. Operational check: confirm training job logs show near-linear throughput gains as worker count increases before assuming the setup scaled correctly.