AWS wellarchitected documentation change
Summary
Added new best practice FSIPERF02-BP02 for selecting GPU/accelerated computing for financial AI workloads with instance recommendations
Security assessment
The change focuses on performance optimization and cost management for AI workloads without mentioning security controls, vulnerabilities, or compliance requirements. Content covers technical specifications for GPU instances and workload optimizations without security context
Diff
diff --git a/wellarchitected/latest/financial-services-industry-lens/fsiperf02.md b/wellarchitected/latest/financial-services-industry-lens/fsiperf02.md index ceedb534c..d47b7d054 100644 --- a//wellarchitected/latest/financial-services-industry-lens/fsiperf02.md +++ b//wellarchitected/latest/financial-services-industry-lens/fsiperf02.md @@ -5 +5 @@ -FSIPERF02-BP01 Select your compute architecture based on workload requirements +FSIPERF02-BP01 Select your compute architecture based on workload requirementsFSIPERF02-BP02 Select appropriate GPU and accelerated computing for AI workloads @@ -20,0 +21,45 @@ The [Financial Services Grid Computing on AWS](https://docs.aws.amazon.com/white +## FSIPERF02-BP02 Select appropriate GPU and accelerated computing for AI workloads + +Financial services AI workloads require careful selection of compute infrastructure to balance performance, cost, and regulatory requirements. Different AI use cases within financial services have varying compute requirements that should guide infrastructure selection. + +**GPU instance selection:** + + * P4d instances for large-scale model training and fine-tuning of foundation models on financial datasets + + * P5 instances for the most demanding AI training workloads requiring maximum GPU performance + + * G5 instances for real-time AI inference workloads like fraud detection and trading algorithm + + * G4dn instances for cost-effective AI inference at scale for applications like document processing and customer service chatbots + + * Inf2 instances powered by AWS Inferentia2 chips for high-throughput, low-cost inference of transformer models + + + + +**Implementation considerations:** + + * Use Amazon EC2 Spot Instances for non-critical AI training workloads to reduce costs. + + * Use Elastic Fabric Adapter (EFA) for distributed AI training across multiple instances. + + * Consider AWS Batch for managing AI training jobs that can run on mixed instance types. + + * Use Amazon SageMaker AI managed infrastructure for production AI workloads with built-in optimization. + + + + +**Accelerated computing for specific financial AI workloads:** + + * Use G5 instances with GPU memory optimized for low-latency inference. + + * Use parallel computing capabilities of P4d instances for risk modeling and Monte Carlo simulations. + + * For document processing and regulatory compliance, use Inf2 instances for transformer-based document analysis. + + * Consider F1 instances with FPGAs for ultra-low latency requirements and algorithmic trading. + + + +