Amazon has launched a new data center architecture that enhances efficiency and reduces energy consumption, employing a quasi-random topology.
Amazon has introduced a groundbreaking data center network architecture aimed at enhancing data transfer efficiency while minimizing energy usage. Dubbed the RNG architecture, this new design employs a "quasi-random" topology, a concept AWS has been developing since 2023, with plans to integrate it into its infrastructure by late 2025.
The centerpiece of this project is the ShuffleBox, a custom optical device crafted by Amazon to rearrange connections between routers dynamically. According to Matt Reder, Vice President of AWS Network Engineering, this innovative approach has enabled the company to streamline its network, effectively eliminating the bottlenecks associated with conventional setups.
In April, the company published a research paper outlining their findings, which reveal significant improvements over traditional data center networks. The RNG architecture is reported to require 69% fewer routers and switches, boost bandwidth by 33%, cut energy consumption by 40%, and lower operational expenses by 27%.
The first implementation of this architecture occurred in Dublin in 2024, followed by expansions into Germany and Spain. Presently, most new AWS data centers are being constructed utilizing this RNG model.
Since the 1980s, data centers have primarily relied on the fat-tree topology, which, while dependable, is inflexible and necessitates complicated cabling. Amazon's network encompasses around 20 million kilometers of fiber optic cables, showcasing its expansive infrastructure.
The RNG concept draws inspiration from a 2012 proposal by University of Illinois researchers suggesting a random graph topology as a better alternative to the fat-tree design. While initial tests with a regular pattern showed promise, they lacked resilience and efficiency, prompting a shift towards the current quasi-random model.
It's important to note that AWS clarified this architecture was not specifically tailored for generative AI applications, as the centralized nature of model training does not lend itself to random graph structures. Instead, the architecture focuses on enhancing the foundational network infrastructure.
Additionally, the surge in data center development has led to rising costs, with PJM Interconnection reporting an increase of $23.1 billion in expenses within its operational area due to this growth.
