A company hosts a custom PyTorch CNN image classifier on a SageMaker endpoint and wants higher throughput and lower latency for users. Which option is the most cost-effective way to improve performance?

Enabling Amazon Elastic Inference (EI) on the SageMaker hosted endpoint is the most cost-effective solution. EI allows you to attach GPU-powered acceleration to existing CPU instances, providing a fractional GPU for inference. This significantly improves throughput and lowers latency for deep learning inference without the higher cost of full GPU instances. Retraining the CNN with more layers (regardless of dataset size) would likely increase model complexity and inference time, worsening performance rather than improving it. Switching to an instance type with multiple GPUs would offer performance gains but is generally much more expensive than using Elastic Inference, making it less cost-effective.