Arista Networks is widening its challenge to Nvidia’s proprietary AI networking stack, rolling out new Ethernet reference architectures and a set of integration partnerships designed to make standard Ethernet a practical foundation for large-scale AI clusters. The company is positioning Ethernet as the scale-up and scale-out fabric of choice for artificial intelligence workloads, contrasting its approach with Nvidia’s NVLink and InfiniBand technologies.
Key facts at a glance
- Arista introduced Ethernet reference architectures for AI scale-up and scale-out deployments.
- The company announced integration partnerships with AMD, Arm, Broadcom, Meta, Microsoft, and Qualcomm.
- Arista is offering 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.
- The designs combine Arista hardware and software with technologies from partner vendors.
- Scale-up is based on ESUN Ethernet, while scale-out uses Ethernet with Ultra Ethernet Consortium features.
- Arista’s Etherlink SU-144 architecture includes orthogonal chassis, cabled backplane, and cross-rack designs.
- The cross-rack design can support up to 1,024 accelerators while maintaining high-bandwidth, low-latency connectivity.
- Arista says it is not selling integrated rack SKUs; it is selling switches and providing software and technology for joint development designs with partners, VARs, and system integrators.
Why Ethernet is central to Arista’s AI strategy
Arista sees a major market opening as AI infrastructure moves beyond Nvidia-only deployments. For years, Nvidia’s proprietary scale-up networking was the primary technology used to build high-performance AI fabrics. But as the total addressable market for scale-up networking expands, and as more non-Nvidia accelerators reach the market, Arista argues that customers want an open, Ethernet-based alternative that avoids locking them into a single vendor’s interconnect stack.
The company’s pitch is that Ethernet can now deliver the high bandwidth, low latency, and reliable in-order delivery required for collective communication in AI systems. Arista executives say the industry no longer needs to treat proprietary interconnects as the default path for scale-up AI networking. Instead, they argue, standard Ethernet can support the performance requirements of modern AI clusters while preserving flexibility across accelerators, switching silicon, optics, physical rack designs, and network software.
Arista’s positioning is not limited to a single product. The company is assembling a broader ecosystem around Ethernet for AI, including reference designs, partner integrations, diagnostics software, and optical technologies. The goal is to give large enterprises and hyperscale customers a path to integrate compute, networking, power, and liquid cooling across increasingly dense AI deployments.
ESUN and UEC provide the standards foundation
The Ethernet for Scale-Up Networking initiative, known as ESUN, was formed by the nonprofit Open Compute Project. Its mission is to advance Ethernet 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 needed for collective communication. That combination is critical for AI training and inference workloads, which depend on tightly coupled accelerators exchanging data at high speed.
The Ultra Ethernet Consortium, or UEC, provides another standards pillar. UEC members include 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 is using UEC features for scale-out designs, while ESUN underpins its scale-up approach.
Partner ecosystem targets non-Nvidia accelerators
Arista’s integration partnerships with AMD, Arm, Broadcom, Meta, Microsoft, and Qualcomm are intended to help customers build Ethernet-based AI fabrics around a wider range of accelerators. Arista executives point to Microsoft Maia, Meta’s MTIA training and inference accelerator, AMD’s Helios, and Google’s Tensor Processing Units as examples of technologies that create opportunities for Ethernet networking vendors.
The message is that AI infrastructure is diversifying. Hyperscalers and large enterprises are increasingly exploring custom silicon and non-Nvidia accelerators to optimize cost, power, and workload fit. Those environments need networking that can support multiple accelerator types without forcing customers into a proprietary fabric. Arista wants to position Ethernet as that common foundation.
Arista will offer its 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. Those components are combined with software and other technologies from partner vendors to create reference architectures rather than fully integrated rack products sold directly by Arista.
Three Etherlink SU-144 reference designs
Arista’s primary scale-up architecture is called Etherlink SU-144, an Ethernet ESUN-based design. It comes in three reference configurations. The first is an orthogonal chassis that 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.
The second is a cabled backplane design that combines a high-density cabled backplane with a modular rack structure. Arista says this approach enables a highly serviceable design, which matters as AI racks become more complex and operators need to maintain or replace components without disrupting entire clusters.
The third is a cross-rack architecture that provides maximum flexibility to scale compute and networking across multi-rack deployments. This design extends the scale-up domain beyond the physical limits of a single rack and can support up to 1,024 accelerators while maintaining high-bandwidth, low-latency connectivity. That ability to scale across racks is important as AI models grow and clusters become denser.
Scale-out features and optical bandwidth gains
For scale-out, 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 features are aimed at handling AI token generation efficiency and the demanding workloads that run across large clusters.
Arista executives say customer designs vary by accelerator power, GPU count, bandwidth, thermal envelope, and rack configuration. The company is emphasizing its capabilities in chassis engineering, thermal design, and signal integrity as differentiators. One rack configuration features bandwidth that scales 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.
Arista says a Tomahawk 7, 200G-based generation could support 6.5 petabytes per second in one rack 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. Arista says this densification translates to nearly a 50% reduction in the overall data center footprint, resulting in significant time and financial savings.
A partner-led rack integration model
Arista is taking a deliberately partner-led approach to rack integration. The company says customers will not find a rack SKU on Arista’s price list because Arista is not selling integrated racks. Instead, it is selling the switches that go into liquid-cooled racks and offering software and technology for integration with custom joint development designs.
Arista is partnering with VARs and system integrators that customers can work with to deliver fully integrated racks. It is also working with different optics vendors as well as partners such as Foxconn, Quanta, and Hive. This model lets Arista focus on switching, software, and validation while relying on partners for rack-level assembly and integration.
The approach 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 rather than as separate components.
NetDI provides diagnostics across environments
The foundation of Arista’s designs is its Network Diagnostics Infrastructure, known as NetDI. This diagnostics and validation layer runs between Arista’s EOS operating system or SONiC, FBOSS, or other network operating systems. 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, and other factors that help cloud companies manage massive scale networks. Arista says 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 racks become denser and more complex, that diagnostics layer becomes increasingly important. Operators need visibility into optics, power, cooling, and switch health across large clusters. Arista is positioning NetDI as a common validation foundation that can work across its own operating system and third-party network operating systems, giving customers a consistent way to manage Ethernet-based AI fabrics as they scale.
Source: Network World News