RAM and VRAM Both Store Data. Here's Why They're Nothing Alike.
Main memory and video memory came from the same place and then split — one obsesses over latency, the other over bandwidth, and HBM ends up bolting memory onto the GPU's roof.
Main memory and video memory came from the same place and then split — one obsesses over latency, the other over bandwidth, and HBM ends up bolting memory onto the GPU's roof.
内存和显存本是同根生,却走上了两条相反的路——一个死磕延迟,一个死磕带宽,到 HBM 干脆把存储架到了 GPU 头顶。
一个被 99% 的人忽视的事实:训练不是瓶颈,推理才是。推理需要内存,海量内存。而这个需求正在创造一个持续十年以上的半导体超级周期。
HBM isn't faster DRAM. It's DRAM rotated 90 degrees — stacked, drilled through with thousands of vertical wires, and bolted next to the GPU on a silicon interposer. That manufacturing nightmare is exactly why three memory makers just crossed a trillion dollars.
一张 H100 超过一半的成本是 HBM。它凭什么这么贵?TSV、硅中介层、垂直堆叠——以及为什么只有 SK 海力士能赚钱。
处理器的速度是指数增长的,内存带宽是线性增长的。这个裂缝是过去二十年计算架构最根本的瓶颈——而你每天都在感受它。
© Xingfan Xia 2024 - 2026 · CC BY-NC 4.0