Chapter 14 completed the hardware context for the entire course. Each GPU generation ships with a different NVLink bandwidth, a different NVSwitch generation inside the node, and the same PCIe constraint for non-SXM form factors -- these three parameters determine the fabric decisions upstream of everything you've learned in Chapters 0 through 13.
NVLink bandwidth has grown from 300 GB/s per GPU (NVLink 2 / V100) to 1.8 TB/s (NVLink 5 / B200 and B300), doubling roughly every two years, with NVLink 6 announced at 3.6 TB/s. NVSwitch silicon has tracked one generation behind, providing enough aggregate switch bandwidth to saturate all GPUs in the node simultaneously. SXM form factor is the prerequisite for full NVLink participation; PCIe form factor GPUs face a hard bandwidth ceiling that only grows with PCIe generation.
GH200's NVLink-C2C eliminates the CPU-GPU PCIe bottleneck by placing both dies on a single package with coherent memory access at 900 GB/s -- a different solution to the memory bandwidth problem than adding more DDR channels. H100 CNX takes the opposite approach, co-packaging GPU and NIC on one module to reduce PCIe hops for InfiniBand-only workloads.
HGX B300 is the platform this course builds against: 288 GB per GPU, 1.8 TB/s of NVLink, and eight 800 Gb/s ConnectX-8 SuperNIC ports per node -- double B200's scale-out attachment, which is the single hardware fact that reshapes leaf sizing, optics selection and congestion tuning in every later chapter. MIG partitions a single GPU into up to 7 isolated instances, with SR-IOV providing the corresponding network isolation via virtual functions on the SuperNIC. Confidential Computing adds hardware-enforced memory encryption that improves tenant isolation at the cost of host observability -- a tradeoff you'll encounter in cloud GPU deployments.