Building the AI Factory

Part 1

Foundations

The AI factory, inside a GPU server, how distributed training talks, and the network math every design depends on.

Part 2

The four networks

Clos, rail-optimized and multi-plane topologies, scale-up networks, optics and racks, the front-end with VXLAN EVPN and BGP, storage networks, out-of-band management.

Part 3

The back-end

RDMA from first principles, where fabrics congest, lossless RoCEv2 (PFC, ECN, DCQCN), hashing and load balancing, Ultra Ethernet, InfiniBand.

Part 4

Operations

Telemetry and troubleshooting, multi-tenancy and security, automation and the fabric as code.

Part 5

Full builds

A 2,048-GPU training cluster, a multi-tenant fine-tuning and inference platform, and distributed inference at the edge.

Appendices

Reference

Glossary, fill-in sizing worksheets, a CCDE Practical decision checklist, and a map from every CCDE AI Infrastructure topic to the chapters that cover it.

For any platform

The principles are the physics, protocols and trade-offs every AI fabric shares. Arista, Cisco, Juniper, NVIDIA and SONiC all implement the same standards (BGP EVPN, PFC, ECN, RoCEv2); configuration examples use Cumulus Linux, and the text says when a feature belongs to one vendor.

For CCDE AI Infrastructure candidates

The book grew out of preparation for the CCDE Practical and its AI Infrastructure area of expertise. CCDE lens boxes show how each topic appears in design scenarios, and Appendix D maps Cisco's public technology list to the chapters, with an honest note on what to study elsewhere. Not affiliated with or endorsed by Cisco.

Analogy first

Every technology starts with a plain-language picture before protocol detail.

Real configurations

Cumulus Linux NVUE examples checked against current documentation, with notes for SONiC and InfiniBand.

Design judgment

Decision logs, worked calculations and review questions, in the style of design certification scenarios.