MLA-C01 Question 268
Select 2You are deploying a machine learning model into a production environment using Amazon SageMaker. The model requires GPU instances for inference to meet latency requirements. However, during development and testing, you aim to minimize costs. What is the most appropriate strategy for provisioning compute resources in both production and testing environments?
- A
Use GPU instances such as
ml.p3.2xlargefor both production and testing to ensure consistent performance. - B
Use CPU instances such as
ml.m5.largeduring testing and switch to GPU instances such asml.p3.2xlargefor production. - C
Use Amazon SageMaker Inference Recommender to evaluate and recommend the best instance type for production.
- D
Use serverless inference in production to automatically scale resources based on traffic, while using CPU instances for testing.
- E
Use AutoML to automatically select the instance type for both testing and production environments.
Show answer and explanation
Correct answers: B, C
Explanation
In production environments, GPU instances are often necessary for low-latency, high-performance inference workloads. During testing, however, CPU instances are cost-effective and sufficient for verifying functionality. Amazon SageMaker Inference Recommender is a valuable tool to evaluate and choose the optimal instance type for production, balancing cost and performance.
- A. Incorrect.
Using GPU instances for both production and testing ensures consistent performance but significantly increases costs during development. This approach is not cost-efficient for testing environments.
- B. Correct.
Using CPU instances for testing minimizes costs while still enabling testing of functionality. Switching to GPU instances for production ensures low latency and sufficient resources for inference.
- C. Correct.
Amazon SageMaker Inference Recommender helps evaluate and select the most suitable instance type for production, ensuring cost-effective and performance-optimized deployment.
- D. Incorrect.
Serverless inference is useful for scaling resources based on traffic patterns, but it is not suitable for scenarios requiring specific GPU hardware for latency-sensitive workloads.
- E. Incorrect.
AutoML automates model training and tuning but does not handle instance selection for testing and production environments.