Cisco Pushes Open Networking as AI's Critical Path at OCP Summit

The company’s message at the event was that the network has moved into the critical path of AI. Cisco Senior Vice President of Engineering Will Eatherton said during a pre-briefing that a large training run a few years ago used hundreds of graphics processing units, while today’s largest clusters use hundreds of thousands. Cisco cited Epoch AI data showing that AI projects that once fit in one or two buildings now span campuses, metros and regions, largely because that is where power and real estate are available.

That shift changes the network’s job. In a traditional data center interconnect, the wide-area network carries many small, independent flows, and a network problem affects a single application or user. In large-scale AI deployments, the network must “scale across” and carry a small number of massive, synchronized flows between accelerators working on the same job. Rakesh Chopra, Cisco Fellow and senior vice president of silicon and systems architecture, said on the call that scale-across puts the long-haul network “directly in line with the workload performance.” If a link slows down, he said, GPUs sit idle, and idle GPUs are the most expensive waste in information technology today.

Cisco’s product news reflects that priority. The Ultra-Dense Small Form Factor Pluggable, or USFP, is described as the first pluggable module designed for 400-gigabit-per-lane signaling, delivering up to 12.8 Tbps and about four times the bandwidth density of today’s leading modules. Cisco’s optics team said the competing XPO approach requires liquid cooling, which no optic in the industry needs today, and creates difficulties in handling, replacing and validating modules. USFP is being standardized as a multisource agreement with TE Connectivity and Amphenol. According to SiliconANGLE, optics is one of the highest cost and supply-chain constraints in AI buildouts, and a multivendor specification should drive higher volumes, greater competition and lower prices.

Cisco also described its approach as “open at every layer of the stack.” Eatherton said this is “not just a SONiC story.” The company said its openness spans P4 programmability at the chip level, Nvidia Reference Architecture compliance in systems, OpenZR+ in optics, a choice of network operating systems including SONiC, FBOSS, a custom NOS, NX-OS or IOS XR, third-party controllers, and the CNCF-governed Cilium ecosystem at the Kubernetes layer.

Eatherton called SONiC a first-class network operating system at Cisco and acknowledged that this is unusual for Cisco in the context of its 20- or 30-year history. SONiC was once viewed inside Cisco as a threat, a way for hyperscalers to commoditize switching, according to the article. Cisco now offers community builds, custom builds using its software development kit, and a turnkey Cisco-validated version under the Nexus One controller. SiliconANGLE said the shift aligns with Chief Product Officer Jeetu Patel’s commitment to make Cisco unrecognizable from what it was two years ago.

The openness extends to Cisco’s Intelligent Collective Networking capability in the G300, which detects remote congestion and link failures and signals back to the ingress leaf in hardware time. Cisco is standardizing that capability through the Internet Engineering Task Force rather than keeping it proprietary. According to SiliconANGLE, five years ago Cisco might not have given away something it considered a competitive advantage.