Cisco Launches 102.4T Silicon One G300 Switch Chip

Cisco Launches 102.4T Silicon One G300 Switch Chip

Cisco has recently unveiled its new Silicon One G300 switch chip, capable of 102.4 terabits per second (Tbps). This advanced technology is designed to meet the ever-growing bandwidth needs of massive AI training and inference clusters.

Key Features of the Silicon One G300 Switch Chip

The Silicon One G300 switch chip boasts 512 serializers/deserializers (SerDes), enabling high-speed connections that can support large-scale deployments. Notably, it allows for the connection of up to 128,000 GPUs with only 750 switches, a significant reduction from the previous requirement of 2,500 switches.

Enhanced Performance and Congestion Management

  • This chip features a collective networking engine with a fully shared packet buffer.
  • It includes a path-based load balancer that enhances link utilization and reduces latency.
  • Compared to traditional packet-spraying methods, Cisco claims a 33% improvement in link utilization and up to a 28% reduction in training time.

Rakesh Chopra, Cisco’s Senior Vice President and Fellow, emphasizes this integration as crucial for handling congestion in AI workflows. The G300’s load-balancing agent continuously monitors traffic, facilitating dynamic network adjustments.

P4 Programmability and Software Adaptability

The Silicon One G300 introduces P4 programmability, allowing users to reprogram their devices for new functionalities without needing hardware replacements. This adaptability is particularly beneficial as technology evolves.

Similar features are found in AMD’s Pensando NICs, which also employ the P4 programming language to quickly adapt to changing specifications.

Applications and Compatibility

Cisco’s G300 chip will be incorporated into its N9000 and Cisco 8000 product lines. Both series will feature 64 1.6 Tbps OSFP cages capable of supporting high-bandwidth applications.

New Optics for Enhanced Efficiency

To complement the G300, Cisco is releasing new 1.6 Tbps pluggable optics. These optics can be divided into eight connections of 200 Gbps each, significantly increasing deployment flexibility.

Additionally, Cisco will introduce 800 Gbps linear pluggable optics (LPO), designed to reduce power consumption by managing signal processing within the G300 itself. This shift promises a substantial power reduction of approximately 30% when used with the N9000 or Cisco 8000 systems.

CPO Technology and Future Developments

While Cisco is not ready to launch copackaged optics (CPO), which could further enhance efficiency, Chopra noted that the company possesses the technology and is exploring opportunities for productization.

Conclusion

With the arrival of the Silicon One G300 chip, Cisco positions itself as a key player in the networking sector, directly competing with Broadcom and Nvidia. Their new technology is expected to ship later this year, coinciding with subsequent generations of rack systems from Nvidia and AMD.