ROOT / COMPUTE // HBM MEMORY MODULES
// Compute · decoded · deep-research pass
HBM Memory Modules, decoded.
High-bandwidth memory, the true bottleneck on every AI GPU.
CURATED INDEX◊ confidence: High
// the signal · what the data says
HBM is the scarcest input in the AI supply chain: three vendors (SK Hynix, Samsung, Micron) supply the entire market and are sold out years ahead. HBM4 (2026) roughly doubles per-stack bandwidth and moves the base logic die onto a foundry process, deepening TSMC's role.
// comparison · every entry, benchmarked & annotated
| Name | Maker | Capacity (GB) | Bandwidth (TB/s) | Stack (hi) | Gen |
|---|
| HBM4 (Rubin) | SK Hynix | 64 | 2 | 16 | HBM4 |
| ◊ Powers NVIDIA Rubin, 2026 |
| HBM4 | Samsung | 64 | 2 | 16 | HBM4 |
| ◊ 1c-nm base die |
| HBM3e 12-Hi | Micron | 36 | 1.28 | 12 | HBM3e |
| ◊ Ships in Blackwell Ultra |
| HBM3e 8-Hi | SK Hynix | 24 | 1.2 | 8 | HBM3e |
| ◊ Blackwell workhorse |
| HBM3 | SK Hynix | 24 | 0.82 | 8 | HBM3 |
| ◊ Hopper H100/H200 era |
// leaderboard · Capacity (GB)
01HBM4 (Rubin)64
02HBM464
03HBM3e 12-Hi36
04HBM3e 8-Hi24
05HBM324
// sources & where to verify
// full index, per-field sourcing & CSV export [ AUTHENTICATE → ]
◊ Deep-research pass, mid-2026. Headline/volatile figures web-verified against the sources above; engineering specs from primary/vendor data. Green = best value in column · confidence: High · verify volatile figures (prices, counts, live feeds) before publishing.