MLS-C01 Question 316
Single answerYou are working on a machine learning application that requires real-time predictions with extremely low latency. Your model is already trained and deployed in Amazon SageMaker. However, you need to optimize the inference process to meet the latency requirements. Which option would best help you achieve this?
- A
Use Amazon SageMaker Multi-Model Endpoints to serve multiple models on the same endpoint.
- B
Configure Amazon SageMaker to use Amazon Elastic Inference (EI) for hardware acceleration.
- C
Switch to using Amazon SageMaker Serverless Inference.
- D
Deploy the model to Amazon SageMaker Hosting Services with an instance type optimized for inference, such as an
ml.inf1instance.
Show answer and explanation
Correct answer: D
Explanation
When ultra-low latency is critical, selecting an instance type optimized for inference, such as the ml.inf1 instance, is the most effective solution. These instances are powered by AWS Inferentia chips, which are purpose-built to accelerate machine learning inference workloads. Other options either do not focus on latency optimization or introduce additional latencies, making them less suitable for this scenario.
- A. Incorrect.
Amazon SageMaker Multi-Model Endpoints are useful for serving multiple models from a single endpoint, which can optimize cost and resource utilization but may increase latency due to model loading time.
- B. Incorrect.
Amazon Elastic Inference (EI) can reduce inference costs by attaching GPU acceleration to CPU instances, but it is not primarily designed to handle ultra-low latency requirements for real-time predictions.
- C. Incorrect.
Amazon SageMaker Serverless Inference is designed for intermittent workloads and may introduce cold start latency, making it unsuitable for scenarios requiring extremely low latency.
- D. Correct.
Deploying the model to Amazon SageMaker Hosting Services with an
ml.inf1instance type, which is optimized for machine learning inference, is the best option for achieving extremely low latency for real-time predictions.