Arista takes Ethernet into the rack, and the network becomes the backplane

Arista Networks Inc., a leading switch vendor, is expanding its AI network strategy to rack scale, signaling a shift in how data centers operate. The company unveiled an open Ethernet rack-scale portfolio that includes both scale-up and scale-out AI fabrics, built with a large ecosystem of partners such as AMD, Arm, Broadcom, d-Matrix, Meta, Microsoft, and Qualcomm.

The lead announcement is Etherlink SU-144, an Ethernet-based scale-up architecture that connects up to 144 accelerators within a single-hop domain and up to 1,024 accelerators across racks. This new architecture is built on the Ethernet for Scale-Up Networking (ESUN) initiative. Arista is aiming to integrate the network as a tightly connected backplane that enhances the value derived from compute resources.

The scale-out component of the portfolio features liquid-cooled rack reference designs in eight, 12, and 16-switch configurations. These are built around Arista's 7060EX7 switches and deliver 102.4 terabits per second, along with various components such as fiber patch panels, cooling manifolds, and power shelves with battery backup. These designs are all supported by the Ultra Ethernet Consortium specifications, Multi-path Reliable Connection resiliency, multi-plane routing, and dynamic and cluster load balancing.

Arista is also introducing next-generation optics, including XPO pluggables and open co-packaged and near-packaged optics, which will increase rack capacity from 1.6 petabits per second to 6.5 petabits, reducing the data center footprint by nearly half. The company is targeting non-Nvidia accelerators, recognizing that heterogeneous compute will become the norm.

Arista's strategy is to provide a single network stack across all three fabrics, including EOS, telemetry, and diagnostics, regardless of the underlying silicon. This network-centric view contrasts with the traditional approach of having separate networks for different silicon types. The company is delivering a reference architecture strategy rather than a SKU launch, shifting the delivery model from a traditional rack assembly to a more integrated and tested design.

However, Nvidia's turnkey stack remains a significant obstacle for Arista, as customers often purchase a complete system without fully considering the Ethernet capabilities. Early adopters of this new architecture are expected to be hyperscalers, neoclouds, and AI infrastructure providers, while most enterprises will likely leverage these architectures through their providers. The success of this strategy will depend on effective coordination among accelerator vendors, rack integrators, and cooling suppliers.