AWS wellarchitected documentation change
Summary
Added guidance for cost optimization in generative AI workloads including model selection, inference optimization, vector storage efficiency, prompt engineering, and agent workflows. Expanded best practices with pricing trade-offs and architecture patterns for AI workloads, and updated serverless architecture examples.
Security assessment
Changes focus exclusively on cost optimization techniques for generative AI workloads. While Amazon Bedrock Guardrails are mentioned, they're discussed in the context of minimizing token waste for cost efficiency, not as a security feature. No vulnerabilities, security incidents, or security-specific enhancements are addressed.
Diff
diff --git a/wellarchitected/latest/financial-services-industry-lens/fsicost11.md b/wellarchitected/latest/financial-services-industry-lens/fsicost11.md index 04167fb45..358299970 100644 --- a//wellarchitected/latest/financial-services-industry-lens/fsicost11.md +++ b//wellarchitected/latest/financial-services-industry-lens/fsicost11.md @@ -13 +13 @@ Cloud cost is an important part of the design and architecture process and is us -Cloud cost is an important part of the design and architecture process and is used in making trade-offs between quality, performance, security and other non-functional requirements. Cloud cost is considered when selecting AWS services (using building block services such as Amazon Elastic Compute Cloud versus using managed services such as Amazon Elastic Container Service). +Cloud cost is an important part of the design and architecture process and is used in making trade- offs between quality, performance, security and other non-functional requirements. Cloud cost @@ -15 +15,22 @@ Cloud cost is an important part of the design and architecture process and is us -Cost factors that go into the selection of cloud resources based on the level of cost optimization provided by pricing models or AWS services include: Savings Plans, Reserved Instances, Amazon EC2 Spot Instances, or Amazon S3 Intelligent-Tiering. Cost trade-offs also include resource-level decisions based on performance (for example, selecting an `XL` instead of a `2XL` resource size). +is considered when selecting AWS services (using building block services such as Amazon Elastic Compute Cloud versus using managed services such as Amazon Elastic Container Service). + +Cloud cost is an important part of the design and architecture process and is used in making trade-offs between quality, performance, security and other non-functional requirements. + +For generative AI workloads, consider the following: + + 1. Model selection based on actual performance requirements against cost + + 2. Inference optimization through batching and caching + + 3. Vector store efficiency and storage optimization + + 4. Prompt engineering for cost efficiency + + 5. Agent workflow cost management + + + + +Cloud cost is considered when selecting AWS services (using building block services such as Amazon EC2 versus using managed services such as Amazon ECS or Amazon Bedrock for generative AI workloads). + +Cost factors that go into the selection of cloud resources based on the level of cost optimization provided by pricing models or AWS services include: Savings Plans, Reserved Instances, Amazon EC2 Spot Instances, or Amazon S3 Intelligent-Tiering. Cost trade-offs also include resource-level decisions based on performance (for example, selecting an XL instead of a 2XL resource size). @@ -20,0 +42,4 @@ The cost of the new workload is measured on an ongoing basis during the workload + * **Pricing trade-offs:** Select foundation or fine-tuned models based on objective price-performance ratios, running periodic evaluation jobs that compare accuracy vs cost. Codify model routing rules (for example, gold, silver, and bronze tiers) to ensure workloads default to cost-efficient models unless premium accuracy is justified. Implement guardrails to cap maximum context length and enforce review or approval for gold-tier model usage. + + * **Architecture patterns:** Introduce serverless RAG orchestrators that automatically short-circuit high-confidence cache hits, reducing duplicate inference calls. Apply response compression or summarization before storage to cut downstream S3 or vector store costs. Use Amazon Bedrock Guardrails and content filters to minimize token waste from rejected or repeated outputs. + @@ -23 +48,6 @@ The cost of the new workload is measured on an ongoing basis during the workload - * **Serverless architecture:** FSI companies often have the need to set up automation for processing events and workflows for technology operations. If you use EC2 instances or databases, you are likely not using 100% of the compute capacity at all times. Many customers only use 10–20% of the available capacity in their EC2 fleet at any point in time. This average is also affected by High Availability and Disaster Recovery requirements, which typically result in idle servers waiting for traffic from failovers. In serverless models such as AWS Lambda or DynamoDB, you pay per-request and by duration of time. Additionally, serverless architectures can lower the overall Total Cost of Ownership (TCO) since many of the networking, security, and DevOps management tasks are included in the cost of the service. + * **Serverless architecture:** FSI companies often have the need to set up automation for processing events and workflows for technology operations. If you use EC2 instances or databases, you are likely not using 100% of the compute capacity at all times. Many customers only use 10–20% of the available capacity in their EC2 fleet at any point in time. This average is also affected by High Availability and Disaster Recovery requirements, which typically result in idle servers waiting + + + + +for traffic from failovers. In serverless models such as AWS Lambda or DynamoDB, you pay per- request and by duration of time. Additionally, serverless architectures can lower the overall Total Cost of Ownership (TCO) since many of the networking, security, and DevOps management tasks are included in the cost of the service. @@ -27 +57 @@ The cost of the new workload is measured on an ongoing basis during the workload - * **Right storage selection:** Select the right storage mechanism to optimize cost across metrics, such as storage, IOPS, and data throughput. You can use a combination of the Amazon S3 family of products or AWS database products such as: Amazon Redshift, Amazon RDS, Amazon FSx, Amazon EBS , or Amazon EFS. For more information about these services, see: [Amazon Storage overview](https://docs.aws.amazon.com/whitepapers/latest/aws-overview/storage-services.html) and [AWS Database.](https://docs.aws.amazon.com/whitepapers/latest/aws-overview/database.html) + * **Right storage selection:** Select the right storage mechanism to optimize cost across metrics, such as storage, IOPS, and data throughput. You can use a combination of the Amazon S3 family of products or AWS database products such as: Amazon Redshift, Amazon RDS, Amazon FSx, Amazon EBS , or Amazon EFS. For more information about these services, see: [Amazon Storage](https://docs.aws.amazon.com/whitepapers/latest/aws-overview/storage-services.html) [overview](https://docs.aws.amazon.com/whitepapers/latest/aws-overview/storage-services.html) and [AWS Database.](https://docs.aws.amazon.com/whitepapers/latest/aws-overview/database.html)