Ben Bajarin· @BenBajarin · X·· 21 天前AI 评分58
AI 导读
Ben Bajarin 转述 AMAT 在投资者问答中关于内存 despec 的评论，并认为其逻辑合理，同时指出许多人也提出过类似观点。评论称训练仍需要 12hi 堆叠，而更多面向推理的专用加速器取决于设计，可能只需 4 或 8 层堆叠，未必需要 12hi。
AI 生成摘要 · 以原文为准
正文
Commentary from $AMAT in investor QA on memory despec seemed logical. Many others have suggested this as well.
Things like training will still require 12hi but more inference specific accelerators (design dependent) may not utilize more than 4 or 8 stacks and thus 12hi is not necessary.
来源：Ben Bajarin · x.com