AI 导读
容量 ≈ 堆叠数 × 每堆叠裸片数 × 每裸片 GB 数 带宽 ≈ 堆叠数 × 接口宽度 × 引脚速率 堆叠高度买的是容量，不是带宽。从 12-Hi 降到 8-Hi 会减少三分之一的 DRAM 裸片——BOM 中供应最紧张的硅片——而带宽持平或略升。每 GB 容量的带宽提升 50%，且恰逢 HBM 涨价之际，$/带宽显著改善。推理受带宽约束：每解码一个 token 都要重新读取权重和 KV cache。(2/2)
正文
Capacity ≈ stacks × dies per stack × GB per die
Bandwidth ≈ stacks × interface width × pin speed
Stack height buys capacity, not bandwidth. Cutting 12-Hi to 8-Hi drops a third of the DRAM dies, the most supply-constrained silicon on the BOM, while bandwidth holds or ticks up. Bandwidth per GB of capacity rises 50%, and $/bandwidth gets meaningfully better right as HBM prices hike. Inference is bandwidth bound: every decoded token re-reads weights and KV cache. (2/2)
来源：@SemiAnalysis_ · x.com