So far the GPU node has been treated as an indivisible thing: eight GPUs stitched together inside the chassis by NVLink, joined to everything else by InfiniBand or RoCEv2. That outward-facing fabric -- the scale-out fabric -- is the subject of Chapters 3 through 11.
This chapter asks a different question. What do you do when the scale-out fabric is not what is holding you back -- when the constraint is the seam between scale-up and scale-out itself?
The answer is the NVLink Switch System: a way of pushing NVLink past the chassis wall so that up to 32 DGX H100 nodes behave like a single very large GPU. On the rack-scale Grace-Blackwell systems the same idea is taken further still, which Act 5B covers.