Start here

A path through AI networking

AI fabrics are ordinary networks pushed to an extreme: thousands of GPUs that move in lockstep, where the slowest link sets the pace. Read these in order.

1. Why AI networks are different

   Synchronized traffic, tail latency, and why averages lie.

   • Essay Why averages lie: oversubscription in AI fabrics
   • Essay Where the GPU's own network ends: scale-up versus scale-out

2. Building the fabric

   The routed underlay every AI fabric starts from, and the decisions inside it.

   • Explainer BGP unnumbered, explained without the hand-waving
   • Decision Decision: one spine ASN or many?

3. Keeping RDMA happy

   Lossless Ethernet, where congestion forms, and the research behind it.

   • Explainer Lossless RoCE in one picture: ECN, DCQCN and PFC
   • Essay The six places an AI fabric congests, and which fix works where
   • Research note Research note: DCQCN, the paper behind RoCE congestion control

4. Using every path

   Why ECMP struggles with AI traffic, and what comes next.

   • Essay One link fails, every job slows: what ECMP hashing really does
   • Explainer Ultra Ethernet in plain English

5. Do the math yourself

   Small calculators for the numbers that decide designs.

   • ECMP collision odds, incast queue growth, end-to-end oversubscription

6. The whole picture

   Eighteen chapters, from the GPU server to three complete builds.

   • Building the AI Factory, with a free sample