Skip to content

Chapter 19: IP Addressing and Planning · Part 2 of 11

Act 1 -- The Address Families: What Needs an IP and Why

Before touching a single octet, you need to understand what actually needs an address in a modern AI fabric. The list is longer than it looks.

Address Families
Everything That Needs an IP in an AI Fabric
LB
Switch Loopbacks
/32
SRV
Server /32 Host Routes
/32
BMC
OOB BMC Addresses
/24 block, one /32 per device
IB
In-Band Management
/32 per DGX
STO
Storage NIC Addresses
/32 per NIC port
VT
VTEP (VXLAN) Addresses
/32 per leaf (or /32 shared for anycast pair)
P2P
P2P Link Addresses
/31 if numbered
VRF
Tenant VRF Subnets
/24 per tenant
LB
Switch Loopbacks
~50 per BasePOD  |  Prefix: /32
Range / Block
10.10.0.0/24 per pod
Owner
Every switch
Purpose
Router identity, BGP router-ID, VTEP IP, UFM polling target, Prometheus scrape target
WHAT BREAKS IF WRONG
If duplicated: BGP table corruption, traffic to wrong switch, SNMP returning wrong counters
EXAMPLE
leaf-01: 10.10.0.1/32 spine-01: 10.10.0.16/32

In a BGP-unnumbered fabric (Ch15 Act 2), the inter-switch links do not get IPv4 addresses. BGP peers over IPv6 link-local. So the address families that DO need planned space are:

Loopback addresses -- Every switch and every server that participates in BGP routing needs a loopback. The loopback is the switch's identity. It is the address UFM uses for SNMP polling, the address Prometheus uses for telemetry scraping, the BGP router-ID, and the VTEP IP for VXLAN encapsulation. If a loopback changes, everything that references that switch breaks simultaneously.

Server host prefixes -- Each DGX server needs an IP on its in-band management or compute interface. In a flat L3 fabric, each server is typically a /32 host route advertised by its directly connected leaf into BGP. In an EVPN-VXLAN design, each server gets an IP within its tenant VRF subnet.

Management addresses -- Three separate planes, all needing distinct ranges:

  • BMC/IPMI (OOB): one IP per DGX, one per switch management port
  • In-band management: one IP per DGX host management interface (CX7 Slot1/Slot2)
  • UFM/monitoring server: one IP

Storage addresses -- Each DGX has two dual-port CX7 storage NICs (4 addresses total per DGX). The storage appliances each have NVMe-oF target portal IPs.

VXLAN VTEP addresses -- In an EVPN design, each leaf switch needs a VTEP IP. This is typically the same as the loopback or a dedicated loopback2 address.

VLAN IDs and VXLAN VNIs -- Not IP addresses but allocated alongside them. VLANs are 12-bit (1-4094). VNIs are 24-bit (1-16777215). They need a mapping scheme.

Deployment Scale
Address Space Requirements by Cluster Size
Total planned IPs: 126
Across compute, OOB, storage, management, and loopback address families
Compute /32 IPs
64 / 8,192 max
Switch loopbacks
10 / 400 max
BMC OOB IPs
8 / 1,024 max
Storage NIC IPs
32 / 4,096 max
In-band mgmt IPs
8 / 1,024 max
VTEP addresses
4 / 80 max
VLANs consumed
30 / 4,094 max
A one-hour addressing session before deployment prevents weeks of debugging.

The scale of a BasePOD and SuperPOD makes discipline mandatory:

ScopeSwitchesDGX nodesServer prefixesLoopbacks
BasePOD~108-3232-128~50
SuperPOD (1 BasePOD)~1032128~50
SuperPOD (20 BasePODs)~2006402560~900
SuperPOD (140 DGX H100)~50140560~200

A full 140-node H100 SuperPOD is a common deployment unit. You need addresses for 140 x 8 = 1120 GPU network interfaces, 140 BMCs, 140 storage NIC pairs, ~50 switches, and associated management services. Get this wrong once and fixing it in production means coordinated downtime across everything.