CCIE #69931 · AI fabrics, data center, low latency, automation

Network designs, proven in the lab.

Writing and labs on networks that can't afford guesswork: GPU fabrics for AI training and inference, data center, low-latency trading and automation. Every design comes with the lab output and packet captures behind it, so you can check the work.

Spine 1Spine 2Spine 3Spine 4Rail 0Rail 1Rail 2Rail 3NIC 0 · 1 · 2 · 3Server 1NIC 0 · 1 · 2 · 3Server 2NIC 0 · 1 · 2 · 3Server 3NIC 0 · 1 · 2 · 3Server 4NIC n of every GPU server lands on rail leaf n: same-rail traffic never crosses a spine

Two ways in

For engineers

Learn it, then prove it in a lab

Explainers that start with why, labs you can rebuild with the configs, output and packet captures to check your results against, and network tools for subnets, multicast MACs and MTU.

Start here · Labs · Network tools

For certification candidates

Study the blueprint, domain by domain

Two study series that follow the published blueprints for NVIDIA's NCP-AIN and the CCDE AI Infrastructure track: the reasoning behind each topic, the numbers worth knowing and the commands to practise, one part at a time and in order.

NCP-AIN series · CCDE AI series · All certifications

Explore by topic

2 articles

Data center & enterprise networking

BGP, EVPN and fabric design for the networks everything else runs on, built and broken in labs.

Browse the topic →

First articles coming soon

Automation

Configuration as code, verification and the tooling that keeps a fabric consistent.

Browse the topic →

Latest articles

All writing →

Coming next

The next articles on the schedule. The certification series publish part by part, in order.

Certification guideComing 23 NovNCP-AIN Spectrum domain: BGP-EVPN multi-tenancy on Spectrum-XWhere EVPN belongs in an AI factory, the underlay and overlay split that lets adaptive routing and tenant VRFs coexist, L3VNIs and symmetric routing, route distinguishers and targets under eBGP, how VXLAN carries DSCP and ECN for RoCE, the NVUE commands to build and verify it, and the difference between routing isolation and performance isolation.
EssayComing 25 NovWhy averages lie: oversubscription in AI fabricsThe airline-overbooking bet behind every oversubscribed network, why it fails for GPU training traffic, and a five-step method for sizing the rest.
Certification guideComing 26 NovNCP-AIN Spectrum domain: telemetry with What Just Happened, NetQ and switch countersThree questions every slow or broken AI fabric raises and the tool that answers each: What Just Happened for drops and their reasons, NetQ for fabric-wide validation including RoCE consistency, and switch counters, histograms and high-frequency telemetry for congestion. Verified commands, the constraints the exam likes, and a triage method.
Certification guideComing 27 NovCCDE AI Infrastructure: from use case to infrastructure, training, fine-tuning and inferenceWhat machine learning, deep learning, LLMs, SLMs and generative AI each ask of compute, memory, network and storage, with the arithmetic a designer needs (16 bytes per parameter, 6ND, the KV cache) and an original scenario where the right answer is not to build a GPU fabric.
Certification guideComing 30 NovNCP-AIN Spectrum domain: DOCA and the SuperNIC on the hostThe host half of the RoCE contract. The DOCA-Host stack and the move from MLNX_OFED to DOCA-OFED, mapping RDMA devices to interfaces, firmware settings and BlueField-3 modes, enabling programmable congestion control, GID tables and RoCE versions, DSCP trust and PFC on the NIC, and how to test a GPU-to-GPU RDMA path.
Certification guideComing 1 DecCCDE AI Infrastructure: placement, data sovereignty, governance and lock-inOn-premises, cloud or distributed; residency versus sovereignty; every data flow an AI system creates; GDPR transfers, the CLOUD Act and the EU AI Act timeline after the 2026 Digital Omnibus; NIST AI RMF and ISO/IEC 42001; moving workloads between sites; and interoperability, with an original healthcare scenario.