Glossary › High Bandwidth Memory (HBM)
What is High Bandwidth Memory (HBM)?
High Bandwidth Memory (HBM) is stacked DRAM placed right next to a processor to feed AI accelerators far more data per second than ordinary memory.
Last reviewed
High Bandwidth Memory (HBM) is a type of DRAM built by stacking several memory dies on top of each other and connecting them with thousands of vertical electrical paths called through-silicon vias (TSVs). The stack sits on the same package as the processor, usually on a silicon interposer, so data travels a very short distance over a very wide interface.
Why it matters for AI
Training and running large AI models is often limited less by raw compute than by how fast data can reach the compute units. A GPU with enormous processing power still waits if memory cannot keep up. HBM solves this with a much wider data bus than conventional graphics or server memory, which is why modern AI accelerators ship with several HBM stacks.
Generations
HBM has moved through successive generations—HBM2, HBM2E, HBM3, HBM3E and HBM4—each raising bandwidth per stack and capacity. Each new generation also raises manufacturing difficulty: stacking more dies, keeping them cool, and achieving good yields are hard, which limits how many companies can supply it at scale.
Who makes it
Only a handful of memory makers produce HBM in volume, and supply agreements with AI chip designers are closely watched as a signal of technology leadership. Because HBM is sold in long-term arrangements and is sensitive to yields, company filings and earnings calls are often the most reliable place to learn about capacity and customers.
What to look for in reports
- Which HBM generation a product uses, and how many stacks.
- Supply announcements and qualification news, ideally confirmed in company filings.
- Capacity expansion plans and packaging (such as advanced packaging lines), which often constrain output.
This glossary entry is general information, not investment advice.