HBM
Memory chips stacked on top of one another and placed right next to the processor, giving AI chips their huge memory speed.
Also: high-bandwidth memory
High-bandwidth memory stacks eight to sixteen layers of DRAM and connects them to the processor through thousands of tiny wires on a shared base, called an interposer. The result is far more bandwidth than ordinary memory. An Nvidia H100 has 80 GB of HBM3 at about 3.35 TB/s, and an H200 has 141 GB of HBM3e at about 4.8 TB/s. Memory size decides how large a model fits on one chip, and memory speed often limits how fast tokens are generated. SK Hynix, Samsung and Micron make almost all HBM, and supply has been tight since 2023.