High-Bandwidth Memory (HBM)
科技
内存
Stacked DRAM wired straight to the GPU — the scarcest commodity in AI.
您正在查看一条技术记录：AI 硬件的一个基础构件——它的功能、谁来制造，以及哪些芯片和系统使用它。
HBM stacks DRAM dies vertically, connects them with thousands of through-silicon vias, and mounts the stack millimeters from the processor on a silicon interposer. The result is terabytes per second of bandwidth that ordinary DIMMs cannot approach — and because model inference is mostly memory-bound, HBM capacity and bandwidth, not FLOPS, usually decide real AI performance. Only SK hynix, Samsung and Micron make it, and supply has been sold out roughly a year ahead throughout the boom.
规格与关键参数
每个数据均标注来源依据：厂商规格、厂商声明、公开报道、估算或官方数据。"未公开披露"即字面含义——我们从不捏造数值。
Overview
Suppliers
SK hynix, Samsung, Micron
Generations
HBM (2015) → HBM2 → HBM2E → HBM3 → HBM3E → HBM4
Why it matters
LLM inference throughput scales with memory bandwidth more than with compute
它如何与整个技术栈相连
本记录的关联链接一览——点击分组可直接跳转。
来源
记录审阅截至 2026-01-15 — 查看 方法论.