Foundations
3 chapters
The machine: nodes, NICs, the operating system and why AI traffic is different
AI & HPC networking · from data hall to die
See the whole cluster.Then every wire in it.
InfiniBand, RoCEv2, congestion control and fabric design, taught on a working model of an HGX B300 cluster. Every section has a 3D model. Every lab runs in a terminal inside the chapter.
- 32
- chapters
- 309
- 3D sections
- 21
- in-chapter labs
01 / 07The cluster
2,048 GPUs on one fabric
Eight scalable units of HGX B300, wired rail-optimised through 64 leaf switches and 32 spines, every link 800 Gb/s. This is the machine the course takes apart.
- 2,048
- GPUs
- 256
- Nodes
- 96
- Switches
The gap
Nobody ever built a Packet Tracer for HPC networking.
Engineers who can debug BGP in their sleep, reason about ECMP by instinct, and design a VXLAN fabric from memory still walk into an AI data centre and find nothing familiar. What they need is scattered across vendor PDFs, conference talks and post-incident writeups.
ProDeploy pulls that material into one open, ordered platform. The chapters explain the hardware and the protocols; the labs let you run commands against live state and watch what happens. Corrections, new labs and better explanations are all welcome.
21
scenario labs in the simulator catalogue
32
chapters published and readable right now
How it works
Read it. See it. Run it.
01 · Read
Chapters with a spine
Hardware, transport behaviour, congestion control and operator workflow, joined into one continuous read instead of eleven browser tabs.
02 · See
A model in every section
Each of the 309 sections opens on an interactive 3D model of the part of the cluster it teaches: the node, the rail, the queue, the optic.
03 · Run
Labs inside the chapter
21 labs run in an emulated terminal right where the idea is taught. As you type, a map of the lab's cluster shows which device answered and which cables the command touched.
dgx-node-a:~$
NIC statistics (ConnectX-7 · eth0):
rx_prio3_pause: 0
tx_discards_phy: 47291
link_state: up
Host-local read
DGX Node A › ConnectX mlx5_0 (eth0) → leaf swp1
Curriculum
Seven parts, from the machine to sign-off.
32 chapters and 21 labs, in the order a cluster is built: the machine, the fabric, the physical layer, scale, then bring-up. Reading costs nothing. Sign in when you want your progress to follow you.
Fabric Operations
6 chapters · 10 labs
The fabric in motion: CLI, InfiniBand, lossless Ethernet, load balancing, topology, NCCL
Physical Layer and Infrastructure
4 chapters · 2 labs
Glass and copper, storage, telemetry and the GPU generations underneath
Scale and Architecture
9 chapters · 2 labs
Scale-up and scale-out: NVLink domains, alternative topologies, routing, packet paths, OOB, addressing, UEC and congestion control
Advanced Networking
2 chapters · 2 labs
Steering and securing traffic: segment routing and AI fabric security
Platform Integration
4 chapters · 5 labs
Spectrum-X in production: RoCE, adaptive routing, EVPN multi-tenancy
Build and Acceptance
4 chapters
Cooling, bring-up, validation and the end-to-end design
Who this is for
Written by a network engineer, for network engineers.
CCNP / CCIE engineers
Reading a BGP table and designing a VXLAN fabric are second nature. InfiniBand and RoCE are not — yet. This is the bridge between the two.
HPC and cluster administrators
The servers are yours already; the fabric underneath them still behaves like a black box. This closes the distance between compute operations and network operations.
Cloud and platform architects
You are drawing GPU infrastructure and need to know exactly what a lossless fabric asks of you, at the protocol level and on the day it breaks.
Network engineers moving into AI infrastructure
AI fabrics are moving fast enough that catching up ad hoc does not work. Get the structured version before the first production incident arrives.
Desktop app
ProDeploy for Mac.
The deployment workspace as a native app: plan the cluster, run the bring-up, and investigate the fabric from your dock.
Download for Mac
v1.14.2 · Apple Silicon · 14 MB
Requires a Mac with Apple Silicon (M1 or later). Ad-hoc signed; not notarized by Apple.
Version 1.10.0 expands the reviewed library to 35 procedures, including fans, memory, batteries, drives, power-shelf controllers, configuration and diagnostics. Exact-model procedures require current official sources; missing documentation, unavailable sources and unverified firmware transitions remain clearly marked. Generate guided procedures without a project and download PDF or Word copies.
Built 2026-10-26T10:20:18.657534Z · source 70c747d869d3131bdb46a89e95727b35632f2af3
SHA-256: ee3f94318e6e9786103084ac2ac314999f71f0f3725e04b3b9d9205bfbfb8007
Install
- 1Download the .dmg and open it.
- 2Drag ProDeploy into Applications.
- 3First launch: right-click the app, choose Open, then Open again. On macOS 15 and later, use System Settings → Privacy & Security → Open Anyway.
Start learning
Learn it in the open.
Reading is free and always will be. Sign in and your progress is saved, on whichever machine you pick it up.