An AI datacenter networking curriculum for enterprise teams
AI infrastructure is being built faster than network teams can retrain for it. GPU clusters put demands on the network that traditional enterprise designs never faced — thousands of synchronized flows, microsecond-sensitive collectives, fabrics where a single congested link stalls a million-dollar training job. Teams that grew up on campus networking need a deliberate path into this world.
This is the curriculum we teach, in the order we teach it, with the reasoning behind each step. Use it as a syllabus for your own program, or run it with us — every module below links to its free written guide.
Design principles
- Build the fabric the way operators do. Underlay first, overlay second, workloads third, observability fourth. Every module ends with a working system the next module extends.
- Real protocols, real state. Everything runs on SONiC + FRR — the open-NOS stack of hyperscale AI fabrics — with learners reading actual BGP tables and kernel forwarding state, not simulator abstractions.
- Traffic or it didn't happen. The point of an AI fabric is collective traffic. The curriculum runs real AllReduce jobs and makes learners watch the network carry them.
Module 1 — The BGP CLOS underlay
What: Bring up a leaf-spine CLOS from bare switches: loopbacks, point-to-point links, eBGP everywhere, ECMP. Why first: Every AI fabric stands on an L3 underlay. Engineers must be fluent in why hyperscalers chose BGP-to-the-leaf (deterministic scale-out, per-link failure isolation, no STP) before anything overlays it. Competence check: All BGP sessions established; every loopback reachable via equal-cost paths through both spines. Guide: Build a BGP CLOS underlay for an AI datacenter fabric
Module 2 — The EVPN-VXLAN overlay
What: VTEPs, VLAN↔VNI mapping, the EVPN address family, Type-2/Type-3 routes, and the Linux kernel dataplane that actually forwards it. Why: Multi-tenant AI clusters need L2 segments stretched across an L3 fabric. EVPN-VXLAN is the industry answer, and its control-plane/dataplane split is where most production incidents live. Competence check: Remote VTEPs learned on every leaf; hosts on the same VNI ping across the fabric. Guide: Build an EVPN-VXLAN overlay on SONiC/FRR
Module 3 — GPU workloads and collective traffic
What: Attach worker nodes to the overlay and run real distributed collectives (AllReduce) across them, then trace the traffic: per-flow ECMP spread, flow-label hashing, what "the job is slow" looks like from the network. Why: This is the module that turns network engineers into AI-infrastructure network engineers. Collectives are the workload; understanding their traffic shape (synchronized many-to-many bursts) explains every design choice in the fabric. Competence check: A multi-worker AllReduce completes; traffic observed balancing across both spines. Guide: Run GPU AllReduce traffic over an EVPN-VXLAN overlay
Module 4 — Streaming telemetry
What: gNMI subscriptions into Prometheus, Grafana dashboards of per-link counters, watching a live training job light the fabric up.
Why: AI fabrics are operated by dashboards, not show commands. SNMP polling can't see a 2-second incast event; streaming telemetry can.
Competence check: Dashboards populate during an AllReduce; learner identifies which links carried which flows.
Guide: Streaming network telemetry with gNMI, Prometheus and Grafana
Module 5 — SRv6 transport
What: A dual-stack IPv6 underlay, then SRv6 uSID locators, kernel End.DT6 endpoints, and headend encapsulation — modern traffic-engineered transport, leaf to leaf. Why: SRv6 is the direction hyperscale transport is moving (source routing, no per-path state in the core). Teams designing next-generation fabrics need hands-on time before it appears in an architecture review. Guides: Dual-stack IPv6 BGP underlay → SRv6 uSID transport
Module 6 — Scale-out architecture
What: Super-spines and multi-pod CLOS — the math and the failure domains of growing beyond a single pod. Why: Whiteboard fluency for capacity planning: when does a pod fill up, what does pod-to-pod oversubscription cost a training job, where do super-spines earn their ports. Guide: Super spines — scaling beyond a single-pod CLOS
Timing and format
Run as a cohort, this is a two-week program at half-day sessions, or four full days intensive. Self-paced, most engineers finish in 15–20 hours of lab time. Mixed works well too: instructor-led for modules 1–3, self-paced for the rest.
Whatever the format, hold the line on two things: every engineer drives their own fabric (shared labs create spectators), and completion is measured by device state, not attendance — BGP sessions up, VTEPs learned, dashboards populated.
Running it with us
This entire curriculum is our AI Datacenter EVPN-VXLAN course plus the SRv6 course, running on live SONiC/FRR fabrics driven from a browser, with machine-graded checkpoints per step. Read every guide free, book a fabric slot to test-drive it yourself, and when you want it for a team — private per-learner fabrics, your pace, instructor-led or self-paced — talk to us.