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。

A former Google TPU engineer who worked on V7 and V8 revealed how TPU actually competes with Nvidia. The answer isn't about chip specs — it's about system-level design, software co-optimization, and a fundamentally different philosophy. Apple, Anthropic, and Meta are all using TPU now. Here's what that means.

听了一期前谷歌 TPU 工程师的访谈，才明白 TPU 和 GPU 压根不是同一道题，硬放一张表里比参数比不出什么名堂。Apple、Anthropic、Meta 都在用 TPU 了，但这东西想用好，门槛比想象的高。