HBM3

The memory generation under H100 and MI300, and the reason the 2023-24 accelerator shortage was as much a memory shortage as a logic one. HBM has to be produced by the memory vendors, SK hynix, Samsung and Micron, then bonded to the compute die by the same handful of advanced-packaging facilities that every large accelerator vendor competes for capacity at. A GPU design can be finished and still be unshippable because the memory supply chain, not the logic foundry, is the constraint.