Amazon unveils 'Resilient Network Graphs' data center network that cuts hardware by 69% and boosts throughput by 33% — now the default for most AWS workloads

Amazon has unveiled a new data center networking architecture that it claims delivers up to 33% higher throughput while reducing network power consumption by 40%, marking what the company describes as the first large-scale deployment of a flat network based on random graph theory, Wired reports. The company revealed that it had been quietly deploying the design in its data centers since last year, confirming that it has already become the default data center network for most AWS workloads.

The architecture, called Resilient Network Graphs (RNG), replaces the hierarchical networking structures that have dominated cloud data centers for decades with a flatter, quasi-random architecture designed to move data more efficiently between servers. Amazon says the design uses 69% fewer networking devices than traditional architectures and can reduce infrastructure costs by up to 45%, potentially translating into billions of dollars in savings across its global cloud footprint.

The company first deployed RNG in a Dublin data center in 2024 before expanding the architecture into facilities in Germany and Spain. AWS says the design is now being rolled out across most newly built data centers and will form the foundation of future deployments.

Data center and AI capability conversations typically focus on processing and storage. However, networking is one of the most important aspects of capability. Every request sent from your device to a cloud application, database, AI model, or storage service depends on data moving rapidly across thousands, or even hundreds of thousands, of servers. As facilities continue to grow, ensuring information reaches the right destination quickly and efficiently has become an increasingly difficult engineering challenge.

For decades, most large-scale data centers have relied on a hierarchical networking structure known as a "fat-tree" topology. In this design, data travels up and down layers of switches and routers arranged in a tree-like hierarchy. While the approach is proven and reliable, it can create bottlenecks and strain available bandwidth. Traffic can become concentrated at specific points in the hierarchy, even when capacity is unused elsewhere in the network. The approach also requires a large number of expensive networking devices.

Researchers have long theorized that a flatter network based on random graph principles could solve many of those issues. Instead of forcing traffic through predefined layers, routers are connected in a highly distributed mesh-like arrangement that creates numerous possible paths between endpoints. In theory, this will improve resilience, increase utilization of available bandwidth, and reduce the likelihood of congestion.

The challenge was making such a network practical. Although random graph networks have been subject to academic research for more than a decade, major obstacles have prevented their successful deployment at hyperscale. Routing traffic efficiently through such a network is considerably more complicated than routing traffic through a predictable hierarchy, and physically connecting millions of fiber-optic links without creating an operational nightmare presents an entirely different challenge.

AWS says it overcame those obstacles with two key innovations: software and custom hardware. The first is a custom routing protocol called Spraypoint, which distributes traffic across large numbers of available paths rather than relying primarily on the shortest path. The second is a passive optical device called ShuffleBox, designed to organize and standardize the immense amount of cabling required to build the network at scale.

According to Amazon, the resulting architecture not only moves data roughly one-third faster than conventional designs but also requires significantly fewer switches and routers. The company says those reductions lower both capital expenditures and operating costs while simultaneously reducing power consumption.

AI models are growing larger and more capable, and so are their user bases. To sustain this growth, hyperscalers have to continue optimizing the entire technology stack. Faster chips remain crucial, but so are advances in cooling, power delivery, storage systems, optical interconnects, and networking. AWS believes RNG is one such advantage. The company rightly considers the achievement of deploying random graph theory at scale, “a breakthrough that will deliver greater reliability and performance for AWS customers, save billions of dollars in hardware, and lower CO2 emissions.”

Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.