Phil Gervasi wrote an interesting article describing Rail-Optimized Networking for AI Training Workloads. Go read it first; I’ll wait.

Does it sound interesting? Were you able to see behind the curtain and figure out what it’s really about?

The way I read it, he described a regular leaf-and-spine fabric in which the workload is placed in such a way that most traffic stays within leaf switches. That has nothing to do with networking; it’s just smart workload placement. Straight from Phil’s article:

A rail isn’t a separate topology or a bypass of the leaf-spine fabric. Instead, it’s a consistent mapping of endpoints to a specific network plane within a shared Clos-based fabric.

I don’t know enough about intra-server forwarding to figure out whether the “stay within the leaf switch and use the server bus to send traffic to the other GPU” idea has anything to do with networking, or is it just using RDMA to write into another part of memory (= application-level solution).

In any case, the idea is not new. This is a slide I had in my Designing a Private Cloud presentation from 2014, and the idea was already pretty old and well known by then.

What about the other idea Phil described: use disconnected leaf switches (rail-only networks)? Back to his article:

Each rail acts as an independent failure domain, which also provides a bounded congestion domain.

Remember the SAN-A/SAN-B networks from the 1990s, using either Fibre Channel or iSCSI?

Long story short: When in doubt, read RFC 1925 ;))