High-bandwidth memory (HBM) is memory interface for SDRAM. It differs from GDDR in that it is stacked 3D, with wider (bus width) and slower silicon-to-silicon interconnects (as opposed to narrow and fast). HBM allows for more power-efficient applications that are integrated on/closer to chips.

HBM is widely used in GPUs, HPC and servers, and machine learning ASICs (Cerebras). Higher bandwidth, lower latency (because of shorter travel distance), better power efficiency, better space efficiency mean it’s used increasingly.