Worth passing on?
Select any line in the story to share just that line. Or keep it for later and find it under Saved.
Read the original at The Register
theregister.com
Fetching the latest from every source…
When one datacenter is no longer enough
Training the largest AI models is no longer simply a question of adding more GPUs.
1 min read
Key points
- Training the largest AI models is no longer simply a question of adding more GPUs.
- Traditional datacenter interconnect was designed to move traffic between sites.
- 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.
The story
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. If one part of the cluster stalls, the impact can ripple across the entire job. In this interview, The Register’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. 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. 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. 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.
