Cisco and Aviz Discuss AI Networking Solutions

Recent discussions at a Cisco and Aviz Networks bootcamp focused on constructing high-performance, lossless networking fabrics for Artificial Intelligence (AI) and large language models (Large Language Model (LLM)). The insights shared are of particular relevance to technical decision-makers overseeing Data Center Operations (DCO) and network architecture.

Cisco 8000 Series and SONiC

Satish Surani, a Director at Cisco, highlighted the capabilities of the Cisco 8000 series, which utilizes Silicon One and the SONiC Operating System (OS) to create a flexible networking environment suitable for demanding AI tasks. Key elements emphasized include Remote Direct Memory Access (DMA) (RDMA) (RDMA over Converged Ethernet (RoCE) v2), Priority Flow Control (Power Factor Correction (PFC)), and Explicit Congestion Notification (ECN), all critical for minimizing latency and ensuring lossless operations in Graphics Processing Unit (GPU) (GPU)-intensive scenarios.

Observability with Aviz ONES

Aviz presented its Open Networking Enterprise Suite (ONES), which offers telemetry for monitoring and enhancing network performance across diverse vendor platforms. The importance of adopting scalable leaf-spine topologies was discussed, with Cisco and Aviz collaborating to provide validated frameworks for deploying large-scale AI infrastructures.

Conclusion

This joint bootcamp provided foundational knowledge on high-performance networking specifically tailored for AI workloads. The techniques and technology outlined are relevant for enterprise IT leaders and decision-makers tasked with optimizing network efficiencies in AI applications. This summary conveys critical updates from the original blog post.