P4 Programmability
Adaptive Packet Processing allows new features to be rolled out post-deployment without silicon spin. One hardware design can be optimized for many roles, reducing unique SKUs and simplifying inventory.
Call For Better Pricing! 844-294-0782 | Free Shipping! 
Cisco Silicon One G300
The next wave of AI innovation
The industry’s most advanced 102.4 Tbps scale-out switching ASIC, purpose-built for AI cluster networks with 1.6T ports, 200 Gbps SerDes, and 252MB fully shared packet buffer.
102.4 Tbps switching for the AI era
Cisco Silicon One G300 is the industry’s most advanced scale-out switching ASIC, delivering 102.4 terabits per second in a single device. With integrated 200 Gbps SerDes, 1.6T Ethernet ports, and a massive 252MB fully shared packet buffer, the G300 enables distributed AI workloads and backend fabrics for massive GPU clusters—delivering measurable improvements in throughput, job completion time, and power efficiency.
Built for AI at every layer
The G300 was engineered from the ground up to optimize total cost of ownership and drive higher profitability for AI network clusters through two strategic pillars: Intelligent Collective Networking and Future-Proofed Infrastructure.
102.4 Tbps
Industry-leading switching capacity contained within a single device, enabling massive-scale AI cluster networking.
1.6T Ethernet Ports
On-chip integrated 200 Gbps SerDes developed in-house at Cisco for low power consumption and high performance with longer reach.
252MB Shared Buffer
Massive fully shared packet buffer embedded directly into silicon, allowing any packet from any port to occupy any available space.
Up to 512 Ports
High radix scaling for a flatter network that connects more compute resources closer to the edge, reducing latency and simplifying topology.
Powering the Cisco N9364F-SG3
The G300 powers the new Cisco N9364F-SG3 switches, offering 64 ports of 1.6T OSFP connectivity in a compact form factor. These switches deliver breakthrough performance for high-density AI clusters in both 100% liquid-cooled and air-cooled deployments.
Fully Shared Packet Buffer
Up to 2.5x increased burst absorption compared to industry alternatives, absorbing synchronized microbursts without packet loss to keep the network running at peak intensity.
Path-Based Load Balancing
Directs traffic across all possible paths and reacts to congestion or faults in hardware at speeds 100,000x faster than software-based tuning, eliminating manual optimization.
Proactive Telemetry
Rich programmable session-level diagnostics help proactively identify and address network faults and optimization opportunities with minimal software intervention at runtime.
Adaptive Packet Processing
P4 programmability enables operators to evolve infrastructure without replacing hardware. One design can be optimized for back-end, front-end, and disaggregated scale-across roles.
Intelligent Collective Networking
Measurable AI infrastructure gains
Intelligent Collective Networking delivers measurable benefits for AI data centers at scale. In simulations, the G300 architecture achieves significant improvements across throughput, job completion time, and power efficiency.
- 33% throughput increase — The larger packet buffer allows higher GPU interconnect traffic without increasing network capacity, building larger networks, or adding more switches, resulting in lower CapEx per deployed GPU.
- 28% JCT reduction — Compared to advanced packet spray implementations, job completion time is significantly reduced, increasing AI compute efficiency and accelerating model convergence.
- 70% greater power efficiency — The high-radix 102.4T liquid-cooled system reduces overall power consumption compared to six equivalent-bandwidth 51.2T air-cooled systems, enabling sustainable AI scaling.
Part of the Cisco Silicon One portfolio
The G300 is the latest innovation in the Cisco Silicon One portfolio, which now powers over 60 Cisco systems. Alongside the P200 for scale-across routing and switching, Silicon One delivers a multigenerational approach to AI networking that prioritizes network efficiency and lower total cost of ownership.
Silicon One Family
G300 (Scale-Out)
102.4 Tbps switching ASIC optimized for AI cluster backend fabrics with 1.6T ports, 252MB shared buffer, and Intelligent Collective Networking.
P200 (Scale-Across)
51.2T deep-buffer routing and switching with quantum-safe encryption, powering Cisco 8000 and Nexus 9000 systems for DCI, universal spine, and core routing.
Nexus 9000 Systems
G300-powered N9364F-SG3 and P200-powered N9364E-SP2R fixed switches deliver both scale-out and scale-across capabilities on the proven N9000 platform.
800G LPO Optics
OSFP 800G Linear Pluggable Optics reduce per-module power by 50% and system power by 30% by integrating signal processing into the Silicon One ASIC.
Scale your AI infrastructure with Silicon One G300
Whether you are building out AI training clusters or scaling inference workloads, our certified Cisco specialists can help you design a Silicon One G300 solution optimized for your requirements.
- 102.4 Tbps with 64 ports of 1.6T OSFP connectivity
- 33% throughput increase and 28% faster job completion
- 70% greater power efficiency with liquid cooling
- Response from a certified Cisco specialist within one business day
Frequently Asked Questions
Cisco Silicon One G300 is a 102.4 Tbps switching ASIC optimized for AI cluster networks. It supports 1.6T Ethernet ports with integrated 200 Gbps SerDes, a 252MB fully shared packet buffer, and high radix scaling for up to 512 ports. It powers the Cisco N9364F-SG3 switches for high-density AI data centers.
The G300 delivers a 33% increase in network throughput through its larger packet buffer, and a 28% reduction in job completion time compared to advanced packet spray implementations. Path-based load balancing reacts 100,000x faster than software-based tuning, and the fully shared 252MB buffer provides 2.5x increased burst absorption to prevent packet loss during synchronized GPU communication.
Adaptive Packet Processing uses P4 programmability to allow operators to evolve their infrastructure without replacing hardware. New features and standards like UEC 1.0 can be implemented post-deployment without a silicon spin. One hardware design can serve multiple roles including back-end AI fabric, front-end networking, and scale-across applications.
The G300 software developer kit (SDK) is available now. Cisco is actively engaged on new customer designs and is targeting the second half of 2026 for first system availability. Contact our team for the latest availability information and to begin planning your deployment.