Arista is intensifying its push to make Ethernet the preferred scale-up networking fabric for artificial intelligence, positioning standard Ethernet as an open alternative to proprietary interconnects such as Nvidia’s NVLink. The company unveiled new Ethernet reference architectures and a set of integration partnerships with AMD, Arm, Broadcom, Meta, Microsoft, Qualcomm, and other industry players. The effort is aimed at large enterprises and hyperscale customers that want to build Ethernet-based scale-up and scale-out designs for increasingly dense AI clusters.

Key Facts at a Glance

• Arista introduced Ethernet reference architectures and integration partnerships with AMD, Arm, Broadcom, Meta, Microsoft, Qualcomm, and others.
• The designs are based on standard Ethernet, with scale-up built on ESUN Ethernet and scale-out built on Ethernet with UEC features.
• Arista is contrasting its approach with Nvidia’s proprietary NVLink and InfiniBand stacks.
• The company will offer 7060EX7 Series switches along with fiber patch panels, integrated liquid-cooling manifolds, drip trays with active leak detection, rack management controllers, and power shelves with backup battery units.
• Arista’s Etherlink SU-144 scale-up architecture includes orthogonal chassis, cabled backplane, and cross-rack designs.
• The cross-rack design supports up to 1,024 accelerators while maintaining high-bandwidth, low-latency connectivity.
• Arista is not selling integrated racks; it is selling switches and software for joint development designs delivered through partners and system integrators.
• NetDI, Arista’s Network Diagnostics Infrastructure, provides hardware-level validation, telemetry, and failure isolation across EOS, SONiC, FBOSS, and other network operating systems.

Arista’s Ethernet Push Targets a Changing AI Networking Market

Arista’s latest move reflects a broader shift in AI infrastructure. For years, Nvidia’s proprietary scale-up networking technologies were the primary option for connecting accelerators inside high-performance AI systems. That limited the choices available to customers and gave Nvidia significant control over the networking layer. As the scale-up market expands and more non-Nvidia accelerators reach production, Arista sees a new opening for Ethernet-based designs that can support multiple accelerator families and give customers more flexibility.

Hardev Singh, vice president and general manager of Arista’s cloud and AI group, said the company’s strategy is built on the idea that customers want an open, standards-based alternative. He noted that Arista’s positioning contrasts with Nvidia’s proprietary NVLink and InfiniBand stacks. According to Singh, the growth of the scale-up total addressable market and the rise of the non-Nvidia accelerator ecosystem create a brand-new opportunity for Arista. The key differentiator, he said, is that the company’s approach is all Ethernet—strictly Ethernet.

Singh pointed to a range of accelerator programs as evidence that the market is broadening. He cited Microsoft Maia, Meta’s MTIA training and inference accelerator, AMD’s Helios, and Google’s Tensor Processing Units. Those technologies, he said, represent opportunities for Ethernet players such as Arista. The implication is that AI networking will not be dominated by a single proprietary fabric, but will instead support a mix of accelerators, switch silicon, optics, and software from multiple suppliers.

ESUN and UEC Provide the Standards Foundation

Arista’s scale-up designs are based on the Ethernet for Scale-Up Networking initiative, known as ESUN. The initiative was formed by the nonprofit Open Compute Project and is intended to advance Ethernet networking technology so it can handle scale-up connectivity across accelerated AI infrastructure. ESUN backers include AMD, Arista, Arm, Broadcom, Cisco, HPE Networking, Marvell, Meta, Microsoft, Nvidia, OpenAI, and Oracle.

According to Arista, ESUN gives network operators flexibility across accelerators, switching silicon, optics, physical rack designs, and network software while retaining the high bandwidth, low latency, and reliable in-order delivery required for collective communication. Those characteristics are essential for AI training and inference workloads, where accelerators must exchange data quickly and consistently across large clusters.

For scale-out networking, Arista is leaning on the Ultra Ethernet Consortium, or UEC. The UEC includes AMD, Arista, Broadcom, Cisco, Eviden, HPE, Intel, Meta, and Microsoft. Its charter is to bring together industry leaders to build a complete Ethernet-based communication stack architecture for high-performance networking. Arista’s scale-out designs incorporate UEC features and are intended to handle the demanding traffic patterns of AI token generation and other accelerated workloads.

Three Scale-Up Designs for High-Density AI Racks

Arista and its partners plan to offer designs that give customers a path to integrate compute, networking, power, and liquid cooling across evolving high-density AI deployments. The three primary designs are part of what Arista calls Etherlink SU-144, an Ethernet ESUN-based scale-up architecture.

Orthogonal Chassis

The orthogonal chassis design uses direct orthogonal connectivity between accelerator and switch blades. It supports accelerator density of 144 XPUs in 100 kW to 400 kW direct liquid cooling envelopes. That design is intended for customers that need very high density within a single chassis and are prepared to use direct liquid cooling to manage the thermal load.

Cabled Backplane

The cabled backplane design combines a high-density cabled backplane with a modular rack structure to enable a highly serviceable design. Serviceability matters in large AI deployments, where components must be replaced or upgraded without disrupting the entire cluster. A modular approach can help operators maintain availability while evolving the system over time.

Cross-Rack Architecture

The cross-rack design is a modular architecture that provides maximum flexibility to scale compute and networking across multi-rack deployments. It enables larger scaling domains across scale-up and scale-out. Bob Laliberte, principal analyst and founder at Liberte Research Group, noted that the cross-rack architecture is particularly noteworthy because it extends the scale-up domain beyond the physical limits of a single rack. It supports up to 1,024 accelerators while maintaining high-bandwidth, low-latency connectivity.

Laliberte also observed that Arista is addressing 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 rather than as separate components.

Scale-Out Designs Focus on Resiliency and Load Balancing

On the scale-out side, Arista is defining UEC-based designs that include Multipath Reliable Connection fabric resiliency, multi-plane routing, intelligent dynamic load balancing, and cluster load balancing. These capabilities are intended to handle AI token generation efficiency and the variable workloads that characterize modern AI applications. As AI clusters grow, the network must be able to route traffic efficiently, recover from failures, and balance load across many paths without introducing bottlenecks.

Singh said customer designs vary by accelerator power, GPU count, bandwidth, thermal envelope, and rack configuration. He positioned Arista’s capabilities in chassis engineering, thermal design, and signal integrity as differentiators. Those engineering disciplines become more important as racks become denser and as liquid cooling moves from an exception to a standard requirement for high-performance AI infrastructure.

Bandwidth Density and Data Center Footprint

Arista said one rack configuration features bandwidth that scales with next-generation interconnects. A 16-switch configuration approaches 1.6 petabytes per second per rack. Other designs using XPO, CPO, and NPO optical technologies could quadruple bandwidth, according to Singh. A Tomahawk 7, 200G-based generation could support 6.5 petabytes per second in one rack while exceeding 100 kilowatts.

Compared with today’s OSFP-based switches, which deliver 1.6 petabytes per second per rack, next-generation optics increase that to 6.5 petabytes per second per rack, according to Singh. That massive densification translates to nearly a 50% reduction in the overall data center footprint, resulting in significant time and financial savings. The implication is that AI operators can pack more networking capacity into fewer racks, which matters as power, cooling, and physical space become constrained.

A Partner-Driven Go-to-Market Model

Arista is taking a partner-driven approach to integrated racks. Singh made clear that customers will not find a rack SKU on Arista’s price list. The company is not selling integrated racks. Instead, it 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 value-added resellers and system integrators that customers can work with to deliver the fully integrated rack. The company is also working with different optics vendors as well as partners such as Foxconn, Quanta, and Hive. That model allows Arista to focus on its core networking and software strengths while relying on partners to handle rack-level integration, manufacturing, and deployment.

The approach may appeal to hyperscalers and large enterprises that want to customize their AI infrastructure rather than buy a fixed, proprietary system. By offering reference architectures and joint development options, Arista can support customers with different accelerator choices, power envelopes, cooling strategies, and rack designs. At the same time, the partner model lets Arista avoid the complexity of selling and supporting complete racks as a single SKU.

NetDI Provides the Diagnostics and Validation Layer

The foundation of all these designs is Arista’s Network Diagnostics Infrastructure, known as NetDI. It is a diagnostics and validation layer that runs between Arista’s EOS operating system or SONiC, FBOSS, or another network operating system. 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 faulty link or optic 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. It also supports rich telemetry for the switch, physical optics, power shelves, and liquid cooling infrastructure.

As AI clusters grow denser and more complex, diagnostics and validation become more important. Operators need visibility into the health of cables, optics, power components, and cooling systems, not just the switch itself. Arista’s message is that Ethernet-based scale-up and scale-out designs can deliver the performance, flexibility, and operational visibility that AI infrastructure requires, while giving customers an alternative to proprietary networking fabrics.

Source: Network World News