Flat Datacenter Networks at Scale at Amazon

Amazon has announced the successful deployment of a new flat, randomized data center network architecture called RNG, which reduces the number of routers needed, enhances throughput, and improves resilience. This development signifies a major technological shift for cloud infrastructure at scale.

Researchers at Amazon, led by principal scientist Giacomo Bernardi and scholar Ratul Mahajan, developed RNG—an architecture inspired by mathematical theories of random graphs and expanders. The design replaces traditional hierarchical fat-tree topologies with a flat, quasi-random network that leverages expansion properties to optimize routing and cabling.

In 2024, Amazon built its first RNG network in a data center near Dublin, Ireland. Follow-up deployments in Germany and Spain in 2025 confirmed the approach’s advantages, including 69% fewer routers, 33% higher throughput, 40% reduction in power consumption, and 27% lower operating costs, compared to conventional fat-trees. By early 2026, RNG became the standard for new Amazon data centers globally.

Impact of RNG on Data Center Scalability and Resilience

The RNG architecture represents a significant advancement in data center networking, enabling more scalable, resilient, and cost-effective infrastructure. Its ability to maintain performance despite router failures and reduce hardware complexity addresses longstanding challenges in large-scale cloud deployments.

This innovation could influence industry standards, prompting other cloud providers and data center operators to reconsider traditional hierarchical designs in favor of more flexible, randomized topologies.

Historical Roots of Random Graphs and Data Center Networking

The concept of optimal routing networks dates back to the 1970s with the study of expanders—graphs with strong connectivity properties. While early research focused on mathematical constructions and theoretical bounds, practical network designs relied on hierarchical fat-trees inspired by Clos topologies since the 1980s.

In 2012, the University of Illinois proposed Jellyfish, a random graph-based approach, but implementation challenges limited its adoption. Amazon’s recent breakthrough builds on decades of research, combining theoretical insights with engineering innovations like Spraypoint forwarding, ShuffleBox cabling, and performance modeling, culminating in the first large-scale practical application of RNG.

“Replacing structured topologies with randomness was a breakthrough, enabling us to build networks that are simpler, more efficient, and more robust.”

— Ratul Mahajan, Amazon Scholar

“The first RNG data center near Dublin proved that this approach is viable at scale, paving the way for widespread adoption.”

— Matt Rehder, VP of Network Engineering

Remaining Questions About RNG Deployment and Performance

While initial results are promising, it remains unclear how RNG networks will perform in diverse, real-world operational conditions at massive scale. Long-term reliability, maintenance complexity, and integration with existing infrastructure are still being evaluated.

Further testing and industry adoption will clarify whether RNG can fully replace or complement traditional topologies in all data center environments.

Next Steps for RNG Technology and Industry Adoption

Amazon plans to expand RNG deployments to additional data centers globally, gathering more operational data. Industry observers expect other cloud providers to evaluate similar approaches, and further research may refine the design. Amazon will also develop tools and standards to facilitate broader adoption and operational management of RNG networks.

Key Questions

What is RNG in Amazon’s data centers?

RNG is a flat, randomized network architecture based on principles of random graphs and expanders, designed to improve scalability, resilience, and efficiency in large-scale data centers.

How does RNG compare to traditional fat-tree networks?

Compared to fat-trees, RNG uses fewer routers, offers higher throughput, consumes less power, and provides more predictable resilience to failures.

Will RNG replace all existing data center networks?

While promising, RNG is currently being tested and adopted gradually. Its suitability depends on operational requirements and scale, and it may complement rather than fully replace traditional architectures initially.

What challenges remain for RNG implementation?

Long-term reliability, operational complexity, and integration with existing systems are still being studied. More real-world deployments will clarify these issues.

Could RNG influence industry standards?

Yes, if proven effective at scale, RNG could lead to new standards for scalable, resilient data center networking architectures across the industry.

Source: Hacker News