Arista's Etherlink SU-144 brings Ethernet scale-up to 144-accelerator domains outside Nvidia's NVLink

Anyone who walks Arista's booth at the OCP Global Summit will spend more time looking at plumbing than at ports. Arista plans to show a fully assembled liquid-cooled rack there, and SiliconANGLE reports the display is built around manifolds, leak handling, fiber management and power more than switching ^[12].

The pitch is an open fabric with a long partner list ^[2]. What a buyer actually commits to on day one is a chassis layout and a cooling plan ^[5]. Vijay Vusirikala, Arista's distinguished lead for AI systems and networks, walked analysts through the three options ^[13].

In the orthogonal chassis, servers sit in front and full-height switch cards in back connect to them directly, with no backplane or cabling ^[6]. It supports 100 to 400 kilowatts of liquid cooling and is meant for training customers who want the biggest scale-up domain ^[6]. The cabled backplane puts the switches in the middle of the rack and reaches servers through copper cable cartridges ^[7]. It gives up density for serviceability. It also lets operators split bandwidth between links inside the rack and links leaving it ^[7].

Cross-rack ties individual accelerator racks together through a dedicated switch rack. Of the three, it is the most flexible and gets to market soonest, since it skips long co-engineering cycles, but it adds interconnect cost and power ^[8]. "Copper runs out of juice very quickly," Vusirikala said ^[9]. Past two racks, the design needs optics ^[9].

The two headline numbers describe different builds. The 1,024 accelerators across racks ^[1] is about 7.1 times the 144 in a single-hop domain ^[14]. Given Vusirikala's two-rack limit for copper, I'd expect any quote at the larger figure to include optics ^[9].

On who this is for, Arista was blunt. Scale-up has tracked Nvidia's NVLink share, and the company said that will not change for Nvidia-based systems anytime soon ^[3]. "There is no near-term opportunity because that is proprietary," Vusirikala said ^[4].

A team on Nvidia hardware keeps NVLink inside the rack. For that team, the relevant part of the launch is scale-out: reference racks of eight, 12 or 16 7060EX7 switches at 102.4 terabits per second each, built and tested by integration partners ^[10]. Sixteen switches at that rate come to 1,638.4 terabits per second ^[15]. That is close to the 1.6 petabits per rack Arista cites today, and the company says next-generation optics raise it to 6.5 petabits while cutting data center footprint by nearly half ^[16].

The open claim extends to software. On the shared Netdi foundation, buyers can choose either Arista's EOS or an open network OS ^[11].

The decision sorts on two questions. The first is whose accelerators go in the rack. On Nvidia, by Arista's own account, scale-up stays on NVLink and the evaluation is about scale-out ^[3]. The second, for everyone else, is which limit binds first. A site that can deliver 100 to 400 kilowatts of liquid cooling and wants the largest training domain fits the orthogonal chassis ^[6]. Where serviceability matters more than density, the cabled backplane fits ^[7]. When the schedule binds first, cross-rack ships fastest and charges for it in interconnect cost and power, with optics required past two racks ^[8][9].