# AWS: Randomized Graph Networks Are Ready for Prime Time — retained record

> Retained news record from news:datacenterknowledge.

Canonical: https://highsignal.app/data/records/6ae5d9869337d577

[← sources](https://highsignal.app/data)


source: [news](https://highsignal.app/data/news) (news:datacenterknowledge)  published 2026-06-01 12:08 ingested 2026-06-02 10:07

[original source ↗](https://www.datacenterknowledge.com/infrastructure/aws-bets-future-cloud-infrastructure-on-flat-network-architecture)

## extract

AWS: Randomized Graph Networks Are Ready for Prime Time
Resilient Network Graphs architecture will become the default network for new non-GPU infrastructure, replacing Clos fabrics.
Amazon Web Services says it has adopted a new data center network design – Resilient Network Graphs (RNG) – as its default architecture for new non-GPU infrastructure, moving away from multi-tier Clos fabrics that have long dominated data center networks.
“This is the new default network for us going forward for any of our non-GPU servers,” said Matt Rehder, vice president of network engineering at AWS, in an interview with Data Center Knowledge.
AWS unveiled RNG this week as a flatter topology built around interconnected switches arranged in a quasi-random graph rather than a strict hierarchical tree. According to the company, the architecture delivers roughly 33% higher throughput than traditional designs, about 40% lower networking power consumption, and a 69% reduction in the number of networking devices across deployments.
The move comes after years of incremental improvements to conventional Clos fabrics. “We were kind of reaching a point of diminishing returns,” Rehder said. “There has to be something else beyond this that can really move the needle.”
From Jellyfish to RNG: Making Random Graphs Work in Production
The concept behind RNG is not new. Researchers have long argued that randomized graph interconnects can use bandwidth more efficiently than hierarchical designs. The 2012 “Jellyfish” paper demonstrated the theoretical advantages, but operational hurdles, such as complex routing decisions and potentially chaotic cabling, kept such topologies in the lab. “If you could build a network like this, it would be better than the Clos networks everyone was building,” Rehder said. "But no one could build it.”
AWS says it bridged the gap with extensive modeling and a production prototype. Early tests validated the concept but exposed practical challenges, most notably cabling complexity. The first version wasn’t pretty. “We did do the effective spaghetti mess of cabling,” Rehder said. “It was extremely onerous.”
Scaling Without Overbuild: Fewer Tiers, Incremental Growth
To make RNG deployable at scale, AWS developed a system called ShuffleBox to organize optical connections while preserving the network’s randomized characteristics. The company also developed a routing protocol, Spraypoint, to manage traffic across the graph. “The modeling has proven out,” Rehder said. “We haven’t had any issues related to that.”
Industry analysts said AWS’s work could have implications beyond a single network deployment. “This announcement marks a massive structural paradigm shift, shattering the 40-year-old rigid Clos/fat-tree blueprint by proving that a flat, randomized network topology can actually operate at hyperscale,” said Sameh Boujelbene, vice president at Dell’Oro Group.
Boujelbene said AWS overcame the physical cabling and routing challenges that historically kept random-graph networking confined to academic research, turning a long-discussed concept into a practical architecture capable of operating at hyperscale.
Ron Westfall, vice president and practice lead for networking and infrastructure at HyperFRAME Research, called the move “an architectural inflection point,” adding that the efficiency, power, and cost benefits AWS cites could prompt other hyperscalers to evaluate similar approaches over time.
RNG Targets Non-GPU Infrastructure
Despite the attention surrounding AI infrastructure, AWS is not positioning RNG as a replacement for the networking used for large GPU clusters. The company continues to use its UltraServer networking architecture for accelerator deployments. “We still think that’s the right network for accelerators specifically,” Rehder said.
Instead, RNG is aimed at the broader cloud infrastructure: compute, storage, databases, and systems increasingly involved in AI inference. “AI is not just a network for GPUs and accelerators,” Rehder said. “It’s a network for GPUs, CPUs, and storage.”
Ultimately, Rehder described the shift as a way to remove networking as a bottleneck while enabling more granular growth. Traditional Clos networks often require operators to deploy extensive networking infrastructure ahead of demand; RNG’s flatter, quasi-random fabric can expand incrementally. “It kind of goes from a rack of network devices – and then you need more, you just add another,” he said. “We never want the network to constrain our business.”

## retained text

No document body was retained for this record — only its metadata was stored.