Arista Networks is widening its push to make Ethernet the foundation of AI scale-up networking, unveiling reference architectures and integration partnerships intended to give large enterprises and hyperscale operators an open alternative to proprietary accelerator interconnects. The company said the new designs combine its 7060EX7 Series switches with fiber patch panels, integrated liquid-cooling manifolds, drip trays with active leak detection, rack management controllers, and power shelves with backup battery units, alongside software and technology from partners including AMD, Arm, Broadcom, Meta, Microsoft, and Qualcomm.

The announcement sharpens Arista's challenge to Nvidia's NVLink and InfiniBand technologies, which have become a common foundation for connecting accelerators inside AI clusters. Arista argues that Ethernet can now deliver the high bandwidth, low latency, and reliable in-order delivery needed for demanding AI workloads, and that an open, standards-based approach gives customers more flexibility as the accelerator market expands beyond a single vendor.

Key facts

• Arista introduced Ethernet reference architectures and integration partnerships with AMD, Arm, Broadcom, Meta, Microsoft, and Qualcomm.
• The package includes 7060EX7 Series switches, fiber patch panels, liquid-cooling manifolds, drip trays with active leak detection, rack management controllers, and power shelves with backup batteries.
• Scale-up designs are based on the Open Compute Project's Ethernet for Scale-Up Networking, or ESUN, initiative; scale-out designs use Ethernet with Ultra Ethernet Consortium, or UEC, features.
• Arista positions the approach against Nvidia's proprietary NVLink and InfiniBand stacks.
• The Etherlink SU-144 architecture comes in three forms: orthogonal chassis, cabled backplane, and cross-rack.
• The cross-rack design can support up to 1,024 accelerators while maintaining high-bandwidth, low-latency connectivity, according to an industry analyst.
• Arista is not selling integrated racks; it is selling switches and software for partners and system integrators to integrate into complete rack designs.
• NetDI, Arista's diagnostics and validation layer, underpins the designs and supports hardware-level validation, signal integrity analysis, secure boot attestation, and telemetry for switches, optics, power shelves, and liquid-cooling infrastructure.

A bet on open Ethernet

At the center of Arista's strategy is the belief that AI networking is moving into a new phase. For years, Nvidia's proprietary scale-up fabric was the primary technology used to build the high-speed domains that connect accelerators into a single logical system. That limited the choices available to customers and tied advanced AI cluster designs closely to one vendor's roadmap. Arista sees a different future, one in which Ethernet becomes the common fabric for both scale-up and scale-out AI infrastructure.

Hardev Singh, vice president and general manager of Arista's cloud and AI group, said the company's position contrasts with Nvidia's proprietary NVLink and InfiniBand stacks. Singh described Arista's approach as strictly Ethernet. The company is betting that as the scale-up market grows and the non-Nvidia accelerator ecosystem expands, customers will want an open networking layer that can support multiple accelerator types and a broader set of hardware suppliers.

That calculation reflects a wider shift in AI infrastructure. Accelerators are no longer limited to one vendor's product line. Microsoft has developed Maia, Meta has introduced its MTIA training and inference accelerator, AMD is advancing Helios, and Google has deployed tensor processing units at scale. Each of those platforms creates an opening for Ethernet switch vendors that can provide a common, flexible connectivity layer, Arista said.

The hardware and rack-level building blocks

Arista's reference architectures are not just switch announcements. They package together the physical and operational elements needed to deploy dense AI clusters. Alongside the 7060EX7 Series switches, Arista is offering fiber patch panels, integrated liquid-cooling manifolds, drip trays with active leak detection, rack management controllers, and power shelves with backup battery units. The goal is to help customers integrate compute, networking, power, and liquid cooling across evolving high-density AI deployments.

These components matter because AI racks are becoming more complex. Higher accelerator power, denser optical connections, and liquid-cooling systems place new demands on power distribution, thermal management, cabling, and serviceability. By providing reference designs that address those elements, Arista is trying to reduce the integration burden on customers and partners while keeping the underlying network based on standard Ethernet.

Three Etherlink SU-144 designs

Arista's primary scale-up offering is called Etherlink SU-144, an Ethernet ESUN-based scale-up architecture. It comes in three designs, each aimed at different deployment priorities.

The first is an orthogonal chassis. It uses direct orthogonal connectivity between accelerator and switch blades, supporting accelerator density of 144 XPUs within 100 kW to 400 kW direct liquid-cooling envelopes. This design is intended for customers that want a tightly integrated, high-density chassis with direct connections between compute and networking elements.

The second is a cabled backplane. It combines a high-density cabled backplane with a modular rack structure to enable a highly serviceable design. The cabled backplane approach gives operators more flexibility in maintenance and component replacement while preserving the density needed for large AI clusters.

The third is a cross-rack architecture. This modular design provides maximum flexibility to scale compute and networking across multi-rack deployments, enabling larger scaling domains across scale-up and scale-out. The cross-rack model is particularly important because it extends the scale-up domain beyond the physical limits of a single rack.

Cross-rack scale-up and the changing unit of deployment

An industry analyst noted that the cross-rack architecture supports up to 1,024 accelerators while maintaining high-bandwidth, low-latency connectivity. That capability reflects a broader infrastructure shift. As AI clusters become denser and move toward liquid cooling, the rack is replacing the individual switch as the practical unit of deployment. Power distribution, thermal management, fiber density, leak detection, and system validation must now be engineered as an integrated solution.

Arista's designs are meant to address that reality. They give customers a path to integrate compute, networking, power, and liquid cooling across high-density AI deployments, while leaving room for different accelerator power levels, GPU counts, bandwidth requirements, thermal envelopes, and rack configurations. Arista said its capabilities in chassis engineering, thermal design, and signal integrity are differentiators in this environment.

Scale-out and bandwidth density

On the scale-out side, Arista is defining UEC-based designs that include Multipath Reliable Connection, or MRC, fabric resiliency, multi-plane routing, intelligent dynamic load balancing, and cluster load balancing. These features are aimed at handling AI token generation efficiency and the demanding workloads that run across large clusters.

Bandwidth density is another central part of the pitch. Singh said one rack configuration can scale with next-generation interconnects. A 16-switch configuration approaches 1.6 petabytes per rack, and designs using XPO, CPO, and NPO optical technologies could quadruple that bandwidth. A Tomahawk 7, 200G-based generation could support 6.5 petabytes in a single rack exceeding 100 kilowatts.

Compared with today's OSFP-based switches, which deliver 1.6 Pb/s per rack, next-generation optics increase that to 6.5 Pb/s per rack. Arista said this massive densification translates to nearly a 50 percent reduction in overall data center footprint, resulting in significant time and financial savings. The claim underscores why networking has become a central design constraint in AI data centers: as accelerators become more powerful, the fabric must carry more data with lower latency and less physical overhead.

Partner ecosystem and go-to-market model

Arista is not trying to sell complete racks. Singh said customers will not find a rack SKU on Arista's price list. The company is selling the switches that go into liquid-cooled racks, along with software and technology for integration with custom joint development designs. Arista is partnering with system integrators, value-added resellers, and other partners that can deliver fully integrated racks to customers.

The partner ecosystem spans accelerator and silicon suppliers, systems builders, and optics vendors. Arista named AMD, Arm, Broadcom, Meta, Microsoft, and Qualcomm among its integration partners, and said it is working with optics vendors as well as partners such as Foxconn, Quanta, and Hive. That model allows Arista to focus on its networking and software strengths while relying on partners to assemble and validate complete rack-level solutions.

The approach also gives customers more choice. Rather than locking into a single vendor's integrated stack, they can combine Arista's Ethernet switches and software with accelerators, optics, power shelves, cooling equipment, and rack designs from a broader supplier base. For hyperscale operators and large enterprises, that flexibility can be important as they try to avoid vendor lock-in and adapt to rapidly changing AI hardware.

NetDI diagnostics foundation

All of the designs rest on Arista's Network Diagnostics Infrastructure, or NetDI, a diagnostics and validation layer that runs between Arista's EOS operating system or third-party network operating systems such as SONiC, FBOSS, or other NOS. Arista's newer 1.6T AI switches integrate NetDI specifically to provide consistent diagnostics across different operating environments.

NetDI can track several switch diagnostics functions, including cable and optic health, device telemetry, troubleshooting, and failure isolation. Those capabilities help cloud companies manage massive scale networks where a single link or optic problem can affect a large AI workload. Arista said NetDI offers deep hardware-level validation, signal integrity analysis, secure boot attestation, and Single Event Upset resiliency, supporting rich telemetry for the switch, physical optics, power shelves, and liquid-cooling infrastructure.

Source:Network World News