Amazon's RNG networking slashes energy and speeds data 2x

Amazon Web Services has quietly deployed a new networking architecture called Resilient Network Graphs (RNG) in its data centers since late last year. The technology combines elements of traditional structured networks with random graph topologies, a concept first explored in academic research like the 2012 Jellyfish project. AWS engineers, including academics recruited for the effort, solved the scalability challenge by designing custom hardware called ShuffleBox that automatically arranges the cabling required for random connections. Matt Rehder, VP of AWS Network Engineering, told WIRED that RNG essentially flattens the network, eliminating bottlenecks that plague conventional fat-tree designs.

Unlike many recent networking advances tied to AI training, RNG is aimed at improving the performance and efficiency of Amazon's everyday cloud infrastructure. Traditional fat-tree topologies use structured layers of switches and routers connected by 20 million kilometers of fiber optic cables globally—enough to reach the moon and back 25 times. RNG reduces complexity and energy use by enabling random interconnections that adapt dynamically. The company published its findings in a paper titled "RNG: Flat Datacenter Networks at Scale." While AI training workloads don't benefit from random graphs due to their coordinated patterns, RNG dramatically improves general cloud traffic, giving AWS a potential competitive edge in data center efficiency.