Amazon says it improved data center networks with a quasi-random design

In short: Amazon says it has started using a new “quasi-random” network design in AWS data centers to move data faster while using less energy.

What happened

Amazon says it recently solved a long-running problem in data center networking, which is how thousands of computers inside a data center pass information back and forth. The company says it has been quietly deploying the design since late 2024.

The approach is described in a paper called “RNG: Flat Datacenter Networks at Scale.” RNG stands for “resilient network graphs,” which Amazon says are neither fully planned like a grid nor fully random. Think of it like planning enough streets to avoid traffic jams, but leaving some connections more flexible so cars have more ways to get around (more possible paths for data).

For decades, many data centers have used a “fat-tree” layout, where data often moves up and down through layers of networking boxes called switches and routers (like traffic lights and intersections for internet data). Amazon says those layers can create bottlenecks and require complicated cabling. AWS vice president Matt Rehder told WIRED the new design “flattens” the network to reduce those bottlenecks.

Amazon also built a new piece of equipment called the ShuffleBox. It is an optical device (it uses light through fiber cables) that helps “shuffle” connections inside the hardware, so the cabling outside can be simpler and neater.

Rehder claims the new design uses 69% fewer routers and switches, delivers 33% higher throughput (more data moved per second), cuts network power use by 40%, and lowers operating costs by 27%. Amazon says it first deployed RNG in Dublin in 2024, then expanded to sites in Germany and Spain, and now uses it in most newly built data centers.

Why it matters

AWS data centers are part of the “cloud,” meaning the behind-the-scenes computers that run many apps and websites. If Amazon can move data faster and use less power, it could help keep online services more responsive and potentially reduce some operating costs, even as demand for computing keeps rising.

Source: Wired