Every GPU generation changes the fabric. Not because the GPU compute architecture changes -- that's CUDA engineering territory -- but because each new generation ships with different interconnect speeds, different form factors, different constraints on how many GPUs you can network together and at what bandwidth. The engineers who built the DGX H100 fabric we've been studying throughout this course made specific choices because of H100's NVLink 4 and SXM5 form factor. Those choices would be different for A100, and they'll be different again for B200.
This chapter is the network-relevant guide to GPU hardware generations. It covers what changes between generations from the fabric engineer's perspective: the NVLink and NVSwitch generation table, SXM vs PCIe form factor implications, the PCIe bandwidth ceiling that limits non-SXM deployments, the Grace Hopper Superchip and what it changes about CPU-GPU communication, the H100 CNX converged accelerator, Multi-Instance GPU partitioning and its SR-IOV network implications, and Confidential Computing -- which introduces a new kind of diagnostic opacity that directly affects how you operate an AI fabric.