A production SageMaker real-time endpoint running the built-in object detection model on a P3 instance shows low GPU utilization. Which deployment change will make better use of provisioned inference resources?

The correct option is to redeploy the model on an M5 instance and attach Amazon Elastic Inference accelerators. This is because Elastic Inference (EI) accelerators are designed to provide cost-effective GPU-powered inference by allowing you to attach just the right amount of GPU acceleration to CPU instances (like M5). This addresses the low GPU utilization on the P3 instance by providing dedicated, scalable GPU resources only when needed, optimizing resource use and cost. Switching to a batch transform job is incorrect because the question specifies a real-time endpoint. Moving to a larger GPU instance like P3dn is incorrect because it would further exacerbate the low GPU utilization issue, as P3dn instances have even more powerful GPUs. Hosting on an Amazon ECS cluster with a P3 instance is incorrect because it still uses the same underutilized P3 instance type, not solving the core problem of inefficient GPU usage.