A large AI job isn’t just an application running over the network. Thousands of GPUs could be working together on the same job, repeatedly exchanging data and synchronizing before they can continue. That makes communication performance part of compute performance. In other words, a slow GPU is a problem, but so is a GPU waiting for data from another GPU.

As these systems have grown, the industry has started describing GPU connectivity in three broad domains known as scale-up, scale-out, and scale-across.

I touched on two of these topics in another recent blog post, but in this article we’ll go deeper into each of these areas.

At first glance, the network inside a modern AI factory can look very similar to a traditional data center network. You’ll see BGP, ECMP, high-radix switches, Clos fabrics, and, in many Ethernet environments, familiar technologies like EVPN-VXLAN.

That may sound a lot like the data center networks we’ve been building for years, and in many ways, it is. The difference is not that AI suddenly replaces traditional data center networking with something completely new. Instead, AI changes what the network is being asked to do. 

In other words, technical differences with design and configuration aside, the real difference to understand is the actual goal of the network. 

EVPN-VXLAN can seem unnecessarily complicated until you stop looking at all the individual protocols and start following a packet through the network. 

IRB, or Integrated Routing and Bridging, allows a leaf switch to provide both layer 2 bridging within a subnet and layer 3 routing between subnets. In an EVPN-VXLAN fabric, there are two primary ways we can do that known as asymmetric and symmetric IRB.

So, to understand symmetric IRB and how it compares to asymmetric IRB, let’s do a packet walk through an EVPN-VXLAN fabric.

One of the most interesting capabilities introduced by EVPN is native active-active multihoming. If you’ve spent time designing data center networks over the last decade, you’ve most likely used or at least encountered technologies like vPC, MLAG, MC-LAG, Virtual Chassis, StackWise Virtual, or other similar vendor-specific ways to dual-home servers and access switches.

These technologies solved a real problem. We wanted to connect a device to multiple switches simultaneously while still allowing all available links to forward traffic. Traditional Layer 2 protocols such as STP prevented loops, but they also left redundant links sitting idle which is a big waste of resources. 

The challenge is that most MLAG implementations are proprietary. They use special peer links and synchronization mechanisms, and usually that means you’ll be locked into a particular vendor architecture.

EVPN Active-Active Multihoming (EVPN-MH) takes a different approach. Instead of relying on proprietary control planes between switches, EVPN uses BGP EVPN itself to advertise multihoming information throughout the fabric. The result is a standards-based solution that provides redundancy, load balancing, faster convergence, and better scalability for modern VXLAN fabrics.

Many modern enterprise and cloud data center networks have standardized on a combination of BGP, EVPN, and VXLAN to build scalable Layer 2 and Layer 3 fabrics. Whether the environment is only a few dozen switches or thousands of leaf and spine devices, EVPN has become the preferred control plane for distributing endpoint reachability information across the fabric.

An important concept to understand when first learning EVPN are the route types used within the EVPN address family. Unlike traditional BGP, which primarily exchanges IP prefixes, EVPN distributes a variety of information. Each route type serves a specific purpose and helps the fabric build a complete picture of where endpoints, gateways, and services exist.

For decades, network architects designed data center networks primarily around application and storage traffic. Whether supporting enterprise applications, web services, virtualization platforms, or cloud-native workloads, the network was the transport mechanism connecting users, applications, and data.

Traffic patterns were characterized by north-south communication flows, predictable east-west application exchanges, and a relatively loose coupling between compute and network performance. High-performance computing environments differed somewhat by emphasizing ultra-high bandwidth, low latency, and efficient inter-node communication for distributed workloads. However, modern AI infrastructure introduces a fundamentally different model.

A good friend of mine is an electrical engineer for a well-known passenger train company in the United States, and we recently talked about how his job, his entire career, has taken a turn from traditional electrical engineering to computer systems integrations and building telemetry pipelines. In other words, as the industry progressed, he wasn’t replaced. Instead, his role evolved.

That probably isn’t a groundbreaking revelation considering almost every industry has been impacted by some form of digitization. Things change, and people adapt. However, what really intrigued me was the new initiative to use AI to augment the train management system.

Hey networking community!

To get you through another week of cutovers, pcaps, and folks blaming the network when it’s yet another DNS issue, I made this song parody just for you 🙂

It’s called Packets, based on one of the best songs to hit the airwaves in February 1996, and in my humblest of opinions, it’s the perfect soundtrack for your next 2am change window…

Enjoy!

(the lyrics are in the post for your viewing convenience)

You can split New York State into several regions, though the exact areas are often hotly debated depending on if you say soda, pop, or know all too well what time the last train leaves Penn Station. When Zig Zsiga and I started the New York Networking Users Group, the idea was to continue with the existing (NY)NUG which had one successful event in Manhattan. But after thinking about it, I thought it would be great to start multiple chapters around the state to accommodate all those crazy different accents (seriously – every region in New York State has a different accent and I love it).

And thus began a brand new chapter of the (US)NUA: the Upstate (NY)NUG.