HBM

High Bandwidth Memory (HBM) is a 3D-stacked DRAM technology that addresses the Memory Wall by placing multiple memory dies vertically atop a logic die, connected through microscopic through-silicon vias (TSVs). This architecture delivers dramatically higher bandwidth than conventional planar DRAM by trading off capacity and latency for parallelism: HBM can sustain hundreds of gigabytes per second across a wide interface while operating at lower clock speeds, reducing power consumption per bit transferred.

HBM is the memory technology behind modern GPU accelerators and is increasingly used in AI training chips, but its cost and manufacturing complexity have kept it from displacing conventional DRAM in general-purpose computing. The technology represents a recognition that the memory wall cannot be solved by faster signaling alone — it requires a fundamental rearchitecting of the memory-processor interface.