How flat is replacing fat in AWS data center networks

2026-05-28 · Source: Amazon Science homepage · Field: Technology & Digital — Cloud Computing & IT Infrastructure, Emerging Technologies & Innovation, Networking · Depth: Advanced, long

Summary

AWS is transitioning its data center networks from traditional "fat-tree" architectures to a more efficient "flat" design, leveraging "quasi-random" network topologies and a novel passive optical component called ShuffleBox. This new approach, named RNG (resilient network graphs), became the default for most new AWS data center builds globally by April 2026. The fat-tree design, while simple, suffers from inefficiency, congestion, and fragility. While purely random networks offer superior routing and resilience, their implementation is impractical due to complex cabling and high computational demands for routing. AWS's solution combines random and deterministic elements, employing the Spraypoint routing algorithm to enable lightweight multipath routing and ShuffleBoxes to simplify physical cabling. This shift results in a 69% reduction in routers, up to 33% better throughput, and a projected 40% decrease in network equipment electricity consumption. Mathematical models, validated with 530 processor-years of simulation, ensure predictable performance.

Key takeaway

For AI Architects and DevOps Engineers designing or optimizing cloud infrastructure, AWS's adoption of RNG flat networks signals a critical shift. You should evaluate how quasi-random topologies and passive optical components like ShuffleBoxes can reduce router count by 69% and improve throughput by 33% in your own data center designs. Consider leveraging similar principles to enhance network resilience and reduce operational costs, ensuring your infrastructure supports demanding workloads transparently.

Key insights

AWS's RNG flat network design, using quasi-random topologies and ShuffleBoxes, significantly improves data center efficiency and resilience.

Principles

• Optimal networks have random topologies for diverse paths and resilience.
• Flat networks reduce router overhead and congestion compared to fat trees.
• Combining random and deterministic elements can overcome practical limitations.

Method

The Spraypoint routing algorithm "sprays" traffic randomly to neighbors, then uses shortest-path routing to designated "waypoints" arranged in rings around destinations, guiding traffic to closer rings.

In practice

• Implement quasi-random topologies for improved network resilience.
• Utilize passive optical components like ShuffleBoxes for practical flat network cabling.
• Develop mathematical models to predict network performance pre-construction.

Topics

• Data Center Networking
• Flat Network Topologies
• Quasi-Random Graphs
• ShuffleBox
• Spraypoint Algorithm
• AWS Infrastructure

Best for: CTO, VP of Engineering/Data, MLOps Engineer, AI Architect, DevOps Engineer, Research Scientist

Related on AIssential

• Navigating uncertainty in Amazon's middle-mile network — Designing for "optionality" rather than perfect conditions builds resilient, adaptive logistics networks.
• The Data Centers Have Arrived at the Edge of the Arctic Circle — Former industrial sites are being repurposed for AI-driven data infrastructure at the Arctic Circle's edge.
• Towards Trustworthy 6G Network Digital Twins: A Framework for Validating Counterfactual What-If Analysis in Edge Computing Resources — A data-driven NDT framework validates 6G cloud-edge what-if analysis for proactive resource scaling.
• Flexible Node Types Are Now Generally Available — Databricks Flexible Node Types ensure cluster launch success by automatically falling back to compatible VM instances.
• AWS Interconnect Reaches General Availability with Managed Multicloud and Last-Mile Connectivity — AWS Interconnect simplifies multicloud and last-mile private network connectivity through managed services and an open specification.
• Presentation: Evolution of a Backend for a Streaming Application — Serverless architectures and strategic patterns enable scalable, resilient, and cost-effective multi-region streaming applications.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Amazon Science homepage.