Skip to content

Chapter 0: How We Got Here -- The Hardware Story · Part 15 of 15

What you now know

You started this chapter knowing enterprise servers and switches. You now know:

  • Why GPU training requires a fundamentally different network
  • How the AllReduce synchronisation barrier creates the lossless requirement
  • The four parallelism types and which ones generate external network traffic vs intra-node NVLink traffic
  • Why RoCEv2 traffic is structurally low-entropy, what flowlets are, and why this makes ECMP fail for AI fabrics
  • Why InfiniBand was adopted from supercomputing, and why RoCEv2 emerged as an alternative
  • The evolution from NIC -> HCA -> DPU and what each adds
  • How AMD Instinct GPUs use Infinity Fabric / xGMI for intra-node scale-up instead of NVLink, and why the external Ethernet fabric stays operationally the same
  • The complete anatomy of a DGX H100 and why each component exists
  • How the three networks of a DGX deployment connect and why they are separate
  • Why the same ConnectX-7 NIC can connect to an InfiniBand fabric or an Ethernet fabric -- and what changes across the entire infrastructure when you make that choice

Everything that follows in this course -- PFC, ECN, congestion control, CLI diagnostics -- is about making the compute fabric work correctly. You now understand what you are protecting and why it matters.

Continue to Chapter 1 ->