what is high bandwidth memory
High bandwidth memory, or HBM, is a type of memory built by stacking multiple memory dies vertically and connecting them with extremely short, wide internal connections, achieving dramatically more data throughput in a compact footprint than traditional flat memory chips — a design specifically aimed at feeding data fast enough to modern high-performance processors and AI accelerators.
```flowchart
{
"rows": [
{ "type": "nodes", "items": [
{ "title": "Traditional memory: flat chips, limited connections", "sub": "data moves through a relatively narrow set of pathways", "tone": "blue" }
]},
{ "type": "arrow" },
{ "type": "group", "title": "HBM stacks memory dies vertically", "items": [
{ "title": "Connected by through-silicon vias", "sub": "many short, wide vertical connections between layers", "tone": "green" }
]},
{ "type": "arrow" },
{ "type": "nodes", "items": [
{ "title": "Dramatically higher bandwidth in less space", "sub": "critical for feeding data-hungry AI accelerators", "tone": "orange" }
]}
]
}
```
**HBM's core innovation is achieving far more bandwidth through sheer width of connection rather than pushing individual connections to run faster.** Traditional memory increases bandwidth mainly by making each data connection run at higher speed, which runs into real physical limits; HBM instead stacks memory dies and connects them with an extremely wide set of short vertical connections running in parallel, achieving high bandwidth through sheer number of simultaneous data paths rather than requiring each individual connection to be pushed to its speed limit.
```svg
```
| Aspect | Traditional memory | High bandwidth memory (HBM) |
|---|---|---|
| Physical structure | Flat, single-layer chips | Multiple stacked dies |
| Bandwidth approach | Fewer, faster individual connections | Many parallel, shorter connections |
| Footprint efficiency | Lower bandwidth per unit of board space | Much higher bandwidth per unit of space |
| Typical use | General-purpose computing memory | AI accelerators, high-performance processors |
**HBM exists largely because feeding data to modern AI accelerators fast enough has become as important a bottleneck as the accelerator's raw compute power itself.** Training and running large AI models requires moving enormous amounts of data between memory and compute constantly; if memory can't supply data fast enough, even the most powerful compute hardware sits idle waiting for it — HBM's dramatically higher bandwidth exists specifically to keep pace with that demand, directly addressing what has become one of the primary bottlenecks in modern AI hardware.
**Manufacturing HBM depends directly on advanced packaging techniques like through-silicon vias, tying its story closely to the broader trend toward stacked-die packaging.** Stacking memory dies and connecting them with the many short vertical connections HBM relies on requires the same TSV-based packaging technology used in other advanced stacked-chip designs — meaning HBM's rise as a technology is closely linked to, and dependent on, the broader maturation of advanced packaging techniques across the chip industry.
**HBM costs meaningfully more to manufacture than traditional memory, so it's used deliberately where its bandwidth advantage clearly justifies that cost.** The added manufacturing complexity of stacking dies and connecting them with through-silicon vias makes HBM significantly more expensive per unit of capacity than conventional memory, which is why it's typically reserved for high-performance applications like AI accelerators and data-center processors, rather than being used in general-purpose computing where its extra bandwidth wouldn't be worth the additional cost.
Read high bandwidth memory through a wide-highway lens: rather than trying to make a single narrow lane of traffic move impossibly fast, HBM builds many parallel lanes stacked close together, moving far more total data even though each individual connection isn't pushed to any extreme — exactly the kind of bandwidth modern AI hardware needs to keep its compute fed.