训练最大规模的AI模型，已不再仅仅是增加更多GPU的问题。随着集群规模的扩大，电力供应正成为一项难以突破的物理限制，迫使基础设施团队思考如何使单个训练任务能够在多个数据中心之间协同运行。这带来了一个截然不同的网络问题。传统的数据中心互连技术旨在实现站点之间的流量传输，而AI训练则提出了更严苛的要求：大规模同步数据流、极低的丢包率，以及GPU之间的紧密协调通信。
Training the largest AI models is no longer simply a question of adding more GPUs. As clusters grow, power availability is becoming a hard physical constraint, forcing infrastructure teams to consider how a single training workload can operate across multiple data centres. That creates a very different networking problem. Traditional datacenter interconnect was designed to move traffic between sites. AI training demands something more exacting: huge, synchronous flows, minimal packet loss and tightly coordinated communication between GPUs.
如果集群中的某个部分出现停滞，其影响可能会波及整个任务流程。在本期访谈中，《The》的Tim Phillips与思科硅架构与系统架构高级副总裁、思科院士Rakesh Chopra探讨了思科所提出的“跨规模扩展”（Scale-Across）需求，以及为何分布式AI基础设施需要一种全新的路由方法。Rakesh解释了AI技术的需求如何正在改变网络的角色——从简单的数据传输通道，转变为计算系统的核心组成部分。
If one part of the cluster stalls, the impact can ripple across the entire job. In this interview, The’s Tim Phillips speaks with Rakesh Chopra, SVP, Silicon and Systems Architecture, Cisco Fellow, about what Cisco calls the “Scale-Across” imperative and why it believes distributed AI infrastructure requires a new approach to routing. Rakesh explains how the demands of AI are changing the role of the network, from simple transport to an integral part of the compute system.
他还讨论了如何让地理上分离的设施表现得更像一台统一的确定性机器，以及在跨长距离光纤链路运行工作负载时，缓冲机制、高速相干光模块和高度集成的硅技术为何至关重要。此次对话还深入探讨了集群内部的结构。Rakesh概述了思科的Silicon One架构和智能集体网络（Intelligent Collective Networking）如何设计用于管理GPU流量的同步爆发式传输、减少瓶颈并缩短任务完成时间。
He discusses the challenge of making geographically separate facilities behave more like a single deterministic machine, and why buffering, high-speed coherent optics and tightly integrated silicon matter when workloads span long-distance fiber links. The conversation also looks inside the cluster itself. Rakesh outlines how Cisco’s Silicon One architecture and Intelligent Collective Networking are designed to manage synchronized bursts of GPU traffic, reduce bottlenecks and improve job completion time.
能效也是另一个核心议题，包括网络功耗与GPU可用能源之间的权衡。访谈还聚焦于安全与系统寿命，涵盖了硬件加速的MACsec和IPsec技术，以及可编程性在帮助基础设施适应不断演变的AI工作负载方面所发挥的作用。
Power efficiency is another central theme, including the trade-off between network consumption and the energy available to GPUs. Security and longevity also come into focus. The interview covers hardware-accelerated MACsec and IPsec, as well as the role of programmability in helping infrastructure adapt as AI workloads continue to evolve.
对于正在规划更大规模、更具分布性的AI环境的基础设施和数据中心领导者而言，本次讨论从实践角度剖析了当单个站点已无法满足需求时会出现的网络挑战。请在下方观看完整采访。
For infrastructure and datacenter leaders planning for larger, more distributed AI environments, the discussion offers a practical view of the networking challenges that emerge once one site is no longer enough. Watch the full interview below.