Reference Architectures¶
"Which architecture should you use to build large-scale AI infrastructure on AWS?" — this section provides validated reference architectures. Each one is a blueprint showing how to combine the individual building blocks (DLAMI, Slurm, containers, EFA, GDS, and more).
The implementation details of each building block are documented as reusable components in the AI Infra Setup, Networking, and Design sections. This section ties those blocks together from the perspective of "here is how this architecture combines them."
Architecture catalog¶
Choose based on your workload scale, operational capability, and orchestration preference.
-
Slurm-based Large-Scale GPU Training Architecture
Combines Capacity Block + DLAMI + Slurm + containers + EFA + GDS to build a distributed training cluster of tens to hundreds of nodes. The simplest and most proven combination.
-
EKS-based (coming soon)
An architecture that runs training and inference on a unified Kubernetes platform. Suited to teams with in-house EKS expertise.
-
AWS ParallelCluster-based (coming soon)
An architecture that automates infrastructure management and Slurm scheduling with ParallelCluster (or PCS).
More architectures are on the way
We currently provide the Slurm-based architecture first. Other architectures — EKS, ParallelCluster, and more — will be added over time. Each reuses the same building blocks (DLAMI, EFA, GDS, and so on); only the orchestration layer differs.