AWS optimizes the entire networking stack for efficient training of generative AI models, offering services like DataSync with Direct Connect for high-throughput data migration, PrivateLink for secure connectivity, EFA and SRD for low-latency network performance, and managed solutions like SageMaker HyperPod.

<div>
<p>This article discusses networking best practices for training and deploying generative AI models on AWS.</p>
<p>Specifically, the article covers:</p>
<ul>
<li>Reference architecture for generative AI on AWS</li>
<li>Data collection methods (online with AWS DataSync, offline with AWS Snow Family, or a combined approach)</li>
<li>Accessing training data (Amazon S3 with gateway endpoints, Amazon FSx for Lustre, AWS PrivateLink)</li>
<li>Optimizing data exchange between training nodes (network topology, OS bypass with Elastic Fabric Adapter, parallelism with Scalable Reliable Datagram)</li>
<li>AWS services for training (EC2 UltraClusters, SageMaker HyperPod)</li>
<li>Conclusion on AWS's optimized networking stack for AI/ML workloads</li>
</ul>
</div>


Networking best practices for generative AI on AWS

Related articles

Related articles

Mar 14
2024
Best practices to build generative AI applications on AWS

Aug 1
2025
Using generative AI for building AWS networks

Sep 29
2025
Build secure network architectures for generative AI applications using AWS services

Jan 31
2024
Generative AI Infrastructure at AWS