# FSIPERF02: How do you select your compute architecture?
<a name="fsiperf02"></a>

## FSIPERF02-BP01 Select your compute architecture based on workload requirements
<a name="fsiperf02-bp01"></a>

 The optimal compute solution for a particular architecture depends on the workload deployment method, degree of automation, usage patterns, and configuration. Third-party solutions can bring their own requirements for infrastructure, which must also be considered. Different compute solutions may be chosen for each step of a process. Selecting the wrong compute solutions for an architecture can lead to lower performance efficiency. 

 Some financial services computing workloads, like risk modeling, are typically loosely coupled and can benefit from event-driven architectures leveraging the scaling capacity of AWS serverless compute options like AWS Lambda and AWS Fargate, combined with messaging services including Amazon SQS and Amazon EventBridge to decouple components. These serverless solutions minimize the overhead of capacity management, automatically scaling in or out to meet demands. 

 Containerized infrastructure can enable financial services institutions to achieve their goals for speed and scalability by providing a standardized environment to leverage across multiple solutions, and supporting the development of microservice-based architectures. Where scale is the primary factor, AWS serverless container compute engine, AWS Fargate, can be used with both Amazon Elastic Container Service (Amazon ECS) and Amazon Elastic Kubernetes Service (Amazon EKS), removing the overhead of managing and provisioning compute resources. 

 For solutions with more specific performance requirements, or needing to run on virtual machines as their compute solution, AWS offers a wide range of Amazon Elastic Compute Cloud (Amazon EC2) instance types, which you can use to select the configuration that is best suited to your needs at any given time. This allows you to both take advantage of the latest CPU technologies as they are released without consideration for prior investment, and choose instance types with features that best suit your workload's requirements, for example instance variants optimized for network, storage, or compute performance. 

 The [Financial Services Grid Computing on AWS](https://docs.aws.amazon.com/whitepapers/latest/financial-services-grid-computing/grid-computing-on-aws.html) whitepaper explores this topic in more detail for specific workloads. 

## FSIPERF02-BP02 Select appropriate GPU and accelerated computing for AI workloads
<a name="fsiperf02-bp02-select-appropriate-gpu-and-accelerated-computing-for-ai-workloads"></a>

 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.