跳到主要内容Skip to main content

HBM(高带宽显存)HBM (High Bandwidth Memory)

HBM (High Bandwidth Memory)

  • 硬件与部署Hardware & deployment
  • 第五站 · 跑起来Station 5 · Running It
  • 引用核验 2026-11-17refs checked 2026-11-17
解决什么问题What problem it solves

堆叠在 GPU 旁边的高带宽显存,决定大模型能不能装下、跑多快。High-bandwidth memory stacked next to the GPU; it decides whether a large model fits and how fast it decodes.

这是什么What it is

HBM 用硅通孔把多层 DRAM 垂直堆叠,再通过宽接口与 GPU 相连。

以 NVIDIA DGX B200 为例:8 张 GPU 共 1,440 GB HBM3e、合计 64 TB/s 带宽,折合单卡约 180 GB 与 8 TB/s。

HBM stacks DRAM layers with through-silicon vias and connects them to the GPU through a wide interface.

NVIDIA DGX B200 example: eight GPUs share 1,440 GB HBM3e and 64 TB/s aggregate bandwidth, about 180 GB and 8 TB/s per GPU.

HBM sits in the GPU package: stacked DRAM layers and a very wide interface deliver far more bandwidth than GDDR, but capacity is capped by the package.
HBM 与 GPU 同封装:多层 DRAM 堆叠、超宽接口,换来远高于 GDDR 的带宽,但容量受封装限制。HBM sits in the GPU package: stacked DRAM layers and a very wide interface deliver far more bandwidth than GDDR, but capacity is capped by the package.
为什么需要知道Why it matters

显存容量决定能装多少参数和 KV cache,带宽决定解码速度;两者常是比算力更紧的瓶颈。Capacity sets how many parameters and KV cache fit; bandwidth sets decoding speed. Either can bind before raw compute does.

别名Chinese aliases

高带宽显存 · HBM3e · HBM 显存

出处Sources
  • NVIDIA · DGX B200 规格NVIDIA DGX B200 specifications ↗
  • NVIDIA · H100 规格NVIDIA H100 specifications ↗