Cisco ships Silicon One G300, unveils G302 chip for AI data centers
Cisco announced volume shipments of its 102.4-Tbps silicon and revealed a new top-of-rack switch to address AI networking bottlenecks.
Key points
- Cisco is shipping the 102.4-terabit Silicon One G300 chip at volume to customers.
- The company unveiled the new 51.2-terabit Silicon One G302 top-of-rack switch.
- These innovations target power and data constraints in artificial intelligence clusters.
Cisco Systems Inc. announced at the Open Compute Project Global Summit that it is shipping its flagship Silicon One G300 silicon at volume, according to SiliconANGLE. The company also introduced the new Silicon One G302 chip for the first time. Both products aim to resolve networking bottlenecks in artificial intelligence data centers by increasing data transfer speeds between graphics processing units and other accelerators.
The Silicon One G300 operates at 102.4 terabits per second and is now available in volume shipments, alongside new Silicon One P200 systems. Cisco stated that these components dramatically increase the bitrate for moving data from one GPU to another. The company also revealed the Silicon One G302, a 51.2-terabit top-of-rack networking switch. This new chip features a 200-gigabit serializer/deserializer interface to convert data streams and support up to 1.6-terabit optics.
Architecture and Design Goals
Cisco described a new “scale-across” architecture designed to link distributed AI data centers into a unified compute fabric. This approach relies on coherent optics and deep-buffered systems to manage the massive data flows required by frontier AI models. The goal is to allow all chips within a cluster to function as a single giant processor. Cisco argued that rebuilding the network as a coherent, programmable architecture is the most effective way to handle these demands.
The Silicon One G302 will ship before the end of the year and supports direct-to-chip liquid cooling. Cisco emphasized that its open infrastructure stack allows hyperscalers, neoclouds, and enterprises to deploy these systems flexibly. The design permits data center operators to swap out components from different vendors, maintaining open flexibility while addressing the raw speed requirements of modern AI workloads.
Impact on Data Center Operations
Networking has become a significant bottleneck for AI data centers as models grow more powerful and require thousands of GPUs. Data center operators face the challenge of scaling these clusters without inflating hardware costs or excessive power consumption. Cisco’s new silicon family addresses this by enabling significantly more data movement from chip to chip. This helps operators maintain efficiency while expanding the size and capability of their AI computing environments.
That suggests a shift toward open, programmable networking stacks as the standard for high-performance AI infrastructure. The introduction of volume-shipped high-bandwidth silicon indicates that the market is ready for these higher-speed interconnects. By supporting diverse vendors and cooling methods, Cisco aims to reduce the complexity of integrating AI accelerators into existing data center fabrics.
What should you do?
Infrastructure teams should review their current data center networking architectures to identify potential bottlenecks in AI workload performance. Evaluate whether your current silicon and optical systems can support the data transfer speeds required by your GPU clusters. Monitor vendor announcements for volume shipments of high-bandwidth chips like the Silicon One G300 and G302 to plan future upgrades.
What to do and how to stay safe: Cisco
- Audit current network throughput capabilities to ensure they meet the data flow requirements of your AI accelerators.
- Assess the flexibility of your current infrastructure to allow component swapping from multiple vendors.
- Plan for thermal management upgrades, as high-bandwidth switches like the G302 may require direct-to-chip liquid cooling.
- Monitor vendor roadmaps for volume availability of next-generation silicon to avoid supply chain delays.
Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
When will the Silicon One G302 chip be available?
Cisco stated that the Silicon One G302 will ship to customers before the end of the year.
What is the data rate of the Silicon One G300?
The Silicon One G300 operates at a speed of 102.4 terabits per second.
Who can deploy these new Cisco networking systems?
The systems are designed for deployment by hyperscalers, neoclouds, and enterprises.
