Cisco Unveils 3nm AI Network Chip Silicon One G300, Defining the Foundation for the "Agent Era" Network with 102.4Tbps Bandwidth

Cisco recently officially released the Silicon One G300, an AI switching chip manufactured using a 3nm process.Designed for large-scale AI cluster backhaul networks, the chip provides a massive 102.4Tbps Ethernet switching capacity per device. It also integrates a 252MB fully shared packet buffer on a single chip, offering 2.5 times the industry's burst traffic absorption capability in parallel computing environments.

The core innovation of the Silicon One G300 lies in its "Intelligent Collective Network" architecture. The fully shared buffer embedded in the chip allows any port to use the entire storage space, preventing packet loss and performance jitter caused by local congestion. Coupled with path-based hardware load-balancing technology, the chip can respond to network failures or congestion events 100,000 times faster than software-based tuning.

Cisco's simulation tests show that this architecture boosts throughput in large-scale GPU interconnect networks by 33%, while reducing AI job completion time (JCT) by 28%.

The G300 continues the P4 programmability lineage of the Silicon One family, enabling operators to extend new protocols and features via software upgrades without replacing hardware. This adaptive packet processing capability allows the same chip to flexibly support various roles, including backhaul, frontend, and data center interconnect (DCI), reducing hardware SKUs and extending the infrastructure lifecycle.

The launch of the G300 is accompanied by the debut of Cisco's new N9100 and 8000 series systems. With a 100% liquid-cooling design, the bandwidth density per system is six times that of previous products, and energy efficiency is improved by nearly 70%. On the optical front, Cisco introduced the 1.6T OSFP module and the 800G linear pluggable optical devices (LPO), which reduce power consumption by 50% compared to traditional retimed modules, leading to a 30% reduction in overall switch power consumption.

The Silicon One G300 Software Development Kit (SDK) is now available, with the first systems featuring the chip expected to be delivered to customers in the second half of 2026. Cisco is also introducing AgenticOps capabilities for data center networks through AI Canvas, enabling a conversational interface that translates complex network diagnostics into actionable solutions.

Cisco stated that the launch of the G300 marks the transition of AI networks from the "Connected Era" to the "Agent Era." With intelligent collective networking and programmable infrastructure as its two core pillars, the company aims to continuously optimize the total cost of ownership (TCO) for AI clusters, providing a lossless, efficient, and evolvable network foundation for generative AI workloads.