MLA-C01 Question 243
Single answerA company has trained a machine learning model to classify images and plans to deploy it for real-time inference. The model requires a GPU for inference due to high computational demand. The company needs a cost-effective solution that scales automatically with fluctuating traffic. Which deployment infrastructure should the company choose?
- A
Deploy the model on Amazon EC2 with GPU instances and use an Auto Scaling group.
- B
Use Amazon SageMaker Hosting Services with multi-model endpoints.
- C
Deploy the model on Amazon SageMaker Hosting Services with GPU instance endpoints.
- D
Use AWS Lambda with Elastic Inference.
Show answer and explanation
Correct answer: C
Explanation
Amazon SageMaker Hosting Services with GPU instance endpoints is the most appropriate solution because it provides a fully managed environment for hosting machine learning models, supports GPU instances for computationally intensive inference, and allows for automatic scaling to handle fluctuating traffic. This reduces operational overhead and provides a cost-effective and scalable deployment solution.
- A. Incorrect.
Deploying the model on EC2 with GPU instances and an Auto Scaling group could work, but managing the infrastructure manually requires substantial operational overhead. This is not the most cost-effective or scalable option compared to managed services like SageMaker.
- B. Incorrect.
Amazon SageMaker multi-model endpoints are designed for hosting multiple models on the same endpoint to save costs. However, they are not ideal for GPU-based inference with high computational demand, as they are better suited for CPU-based use cases.
- C. Correct.
Amazon SageMaker Hosting Services with GPU instance endpoints is the best choice for this scenario. It provides managed infrastructure, supports GPU-based inference, and can automatically scale instances based on traffic, meeting both the cost-effectiveness and scalability requirements.
- D. Incorrect.
AWS Lambda does not support GPU-based inference, even with Elastic Inference. This option would not meet the computational requirements of the model.