AI Networking & High-Speed Optical Cabling

Overview

AI training clusters depend on very high-bandwidth, low-latency networks in which a single marginal link can slow an entire job. This course explains GPU fabric topologies, InfiniBand and lossless Ethernet (RoCE), 400G/800G optics and cables, and how to plan, install and test high-density MPO cabling for AI pods.

Learning Outcomes

On completion of this course, you will be able to:

• Describe GPU cluster fabrics, including rail-optimised leaf-spine designs
• Compare InfiniBand and RoCE-based Ethernet fabrics at a conceptual level
• Select DAC, AEC, AOC or pluggable optics by reach, power and density
• Plan MPO-12/MPO-16 trunking, polarity and loss budgets for 400G/800G links
• Install, inspect and test high-density cabling for AI pods
• Separate compute, storage and management networks in pod layouts

Module Outline

1. Module 1: Why AI changes the network
2. Module 2: GPU servers, NICs and fabric topologies
3. Module 3: InfiniBand and lossless Ethernet (RoCE)
4. Module 4: 400G and 800G optics and form factors
5. Module 5: Copper, active and optical cable choices
6. Module 6: MPO structured cabling, polarity and loss budgets
7. Module 7: Cabling the AI pod: density, pathways and weight
8. Module 8: Inspection, cleaning and testing at scale
9. Module 9: Data centre interconnect for AI

The module outline is indicative and may be refined as materials are updated.