Cisco and Nvidia Push Open Silicon, Rack and Optical Designs for AI Factories at OCP Summit

Cisco said its flagship 102.4-terabit Silicon One G300 chip is now shipping in volume, alongside new Silicon One P200 systems. The company also introduced the Silicon One G302, a 51.2-terabit top-of-rack switch featuring a 200-gigabit serializer/deserializer interface, direct-to-chip liquid cooling and support for up to 1.6-terabit optics. Cisco said the G302 will ship to customers before the end of the year. The products are designed to increase the bitrate at which data moves from one graphics processing unit to another, part of Cisco's scale-across architecture for linking distributed AI data centers into a unified compute fabric.

Cisco and Nvidia are jointly defining a new optical module format. Cisco called it an Ultra-Dense Small Form Factor Repluggable optical module, while Nvidia described it as an Ultra-dense Small Form Factor Pluggable, or USFP. The module is designed to handle 400-gigabit-per-second-per-lane electrical signals, packing 16 full-duplex lanes into a standard transceiver form factor for up to 12.8 terabits per second of total capacity. The companies said it offers about four times the bandwidth density and 1.3 times the power efficiency of existing 200G optics, eliminates the need for complex embedded cooling circuits and uses a two-dimensional array connector based on a two-step insertion design to maximize signal integrity. Nvidia Senior Vice President Gilad Shainer said, "Building the world's largest AI factories starts with rethinking the smallest of connections. With USFP, Nvidia and Cisco are building an ecosystem to reinvent pluggable optics, connecting more computing with greater efficiency to power the next generation of AI factories."

Cisco also introduced the Cisco 8001, an out-of-band network management system powered by its Silicon One A100 chips. The company said it is designed to replace fragmented legacy management tools with a standardized, programmable environment that can support rapid fixes during outages. Cisco announced updates to its networking management software and a new release of the open-source SONiC network operating system. SONiC's Live Protect runtime security feature, available next year, is intended to identify and remediate vulnerabilities in live AI workflows without disruption. Kevin Wollenweber, Cisco's senior vice president and general manager of data center and internet infrastructure, said AI requires a new approach to scaling data center networks because many workloads are now distributed across multiple physical sites. "It's no longer just about individual data centers or clusters, but about how we connect compute at massive scale without compromising performance or resilience," he said.

Nvidia contributed Vera Rubin rack and compute tray designs to the Open Compute Project through its Nvidia MGX architecture, giving system builders common blueprints for AI factories, including rack layouts, networking, power and cooling systems. Through MGX 1.2, Nvidia is also providing open components for liquid cooling, rack-level interconnect infrastructure, modular liquid-cooled fanless I/O bays and standardized power. The racks include the Vera central processing unit rack design and extend the MGX architecture beyond accelerated computing to handle intense agentic AI tasks. Joe DeLaere, AI infrastructure solutions lead at Nvidia, said, "The real bottleneck for the industry right now is how fast the world can manufacture and deploy AI factories."

Nvidia said its Vera CPUs deliver roughly twice the performance of its previous Grace generation on several workload benchmarks and a 1.8-times improvement on Terminal-Bench, which measures practical agent tasks involving code containers. DeLaere said agents need to perform orchestration between model calls, execute code, process data and run evaluations.

Nvidia said Meta Platforms Inc. is adopting its Spectrum-6 Ethernet switch for its next-generation AI factory network, which will connect directly to Vera-Rubin deployments. Spectrum-6 is Nvidia's 102.4-terabit-per-second Ethernet switching chip. Nvidia is also adding support for Google's Falcon Transport to its Spectrum-X Ethernet networking ecosystem, beginning with future ConnectX-10 SuperNICs, with a roadmap that extends to BlueField hardware. Falcon is designed to move traffic over multiple paths, respond to congestion, recover lost packets and maintain secure connections in large-scale networks.