AIF-C01 Question 98
Single answerA company is planning to implement a generative AI application using Amazon Bedrock. They need to balance cost efficiency with high responsiveness for real-time text generation workloads. The application will operate across multiple regions and requires minimal latency. Which approach should they choose to optimize costs while meeting their requirements?
- A
Use a pre-trained foundation model provided by Amazon Bedrock with on-demand pricing and deploy it to multiple regions.
- B
Fine-tune a foundation model with custom data, host it on Amazon Bedrock, and provision it with high throughput reserved capacity.
- C
Use a pre-trained foundation model provided by Amazon Bedrock and deploy it in a single region with auto-scaling enabled.
- D
Deploy a fine-tuned model on Amazon SageMaker hosting with provisioned instances in multiple regions for better control over token-based pricing.
Show answer and explanation
Correct answer: C
Explanation
The best approach is to use a pre-trained foundation model in a single region with auto-scaling enabled. This provides cost efficiency and responsiveness by leveraging Bedrock's pre-trained models without the need for expensive fine-tuning or regional duplication. Auto-scaling ensures that resources adjust dynamically based on demand, minimizing costs during low-usage periods while maintaining performance during high-demand scenarios.
- A. Incorrect.
Using a pre-trained foundation model with on-demand pricing and deploying it to multiple regions will ensure availability but is likely to increase costs significantly due to regional duplication.
- B. Incorrect.
Fine-tuning a foundation model and provisioning it with high throughput reserved capacity could improve performance but would increase costs and may not optimize for real-time responsiveness.
- C. Correct.
Using a pre-trained foundation model in a single region with auto-scaling enabled balances cost efficiency and responsiveness by optimizing resource utilization and scaling only when demand increases.
- D. Incorrect.
Deploying a fine-tuned model on Amazon SageMaker hosting with provisioned instances in multiple regions offers more control but requires higher upfront costs and may not be as cost-efficient for real-time workloads.