AI hardware · Layer
High-bandwidth memory (HBM)
Stacks of DRAM placed right next to the GPU die to feed it data fast enough. AI workloads are often memory-bandwidth bound, so HBM capacity and speed frequently set the practical ceiling on performance.
Key facts
- HBM is built by vertically stacking DRAM dies and connecting them with through-silicon vias, then placing the stack beside the processor on a shared interposer.
- Only three companies make HBM at scale: SK Hynix, Samsung, and Micron.
- Each HBM generation (HBM3, HBM3E, HBM4) raises bandwidth and capacity, and supply is typically sold out generations in advance.
- HBM is one of the tightest links in the whole AI hardware chain: a shortage here throttles finished accelerator output directly.
Where it bottlenecks
HBM yield and capacity are a primary supply limit; qualifying new HBM on an accelerator is slow and a single supplier can constrain the market.
Who dominates it
SK HynixSamsungMicron
Companies in this layer
Micron
United StatesThe third major HBM supplier and a large memory maker.
Samsung
South KoreaMajor HBM and DRAM supplier and an alternative leading-edge foundry.
SK Hynix
South KoreaLeading supplier of high-bandwidth memory for AI accelerators.