HBM

HBM stands for high-bandwidth memory. Multiple DRAM dies are stacked vertically and linked through connections that include through-silicon vias and bonding structures. The wide interface allows many data connections to operate in parallel. (Micron: high-bandwidth memory)

In Micron’s architectural explanation, HBM sits near a CPU or GPU in a system-in-package, connected through a silicon interposer. This is a package-level memory arrangement, not a description of an interchangeable memory module. (Micron: high-bandwidth memory)

Keep the memory and processing roles separate. A GPU performs suitable parallel computation, while HBM provides storage and data transfer for the processing device. Calling an accelerator an “AI chip” does not make its attached memory another processor. (Intel: CPU vs. GPU) (Micron: high-bandwidth memory)

As an arithmetic example, imagine a memory system that can hold 16 GB and sustain a transfer rate of 100 GB/s. The first number describes an amount; the second describes a rate. Moving 10 GB at that sustained rate would take 0.1 seconds if no other cost applied.

These invented numbers explain units; they are not an HBM product specification or a benchmark. When reading actual specifications, keep capacity and bandwidth in separate columns and identify whether each figure describes one stack or a complete system.

Use these questions as a reading checklist, not as a device-selection recommendation. Exact manufacturer documentation takes priority over a general definition.

• Record the exact memory generation and manufacturer part information.
• Separate capacity from bandwidth and note the units.
• Distinguish per-stack figures from aggregate system figures.
• Use package and interface documentation for design decisions.

For the broader explanation, explore hbm bandwidth, capacity and packaging. This page defines the term; the companion guide connects it to the rest of the system.