工程师复盘：megakernels与ASIC已死，GPU调优交给强化学习
一个工程师复盘自己的神预言：megakernels和推理ASIC都没戏，GPU调优会被RL接管，NVIDIA靠软件和HBM赢。
作者在 @swyx 播客上复盘两个月前的争议言论，坚持认为 megakernels 已死。他的理由是内核复杂度太高，没有推理服务商会在生产中用 67k 行手工融合前向传播内核。
他认为 GPU 内核优化是最适合强化学习的任务，AI 代理用 ncu CLI 和 MCP 就能自动调优。他还预测数据中心 Transformer 推理 ASIC 因架构未收敛而失败，NVIDIA 的护城河在于 HBM 容量带宽和软件工具链。
"With megakernels, honestly I'm very bearish. [Megakernels] was a good research direction and it seems very... intuitively and theoretically, it's nice. [However], the kernel complexity itself makes it very difficult to write a very optimized megakernel."
This report is categorized under AITOP's "Industry" desk. Key entities and technical topics under continuous tracking include megakernels, NVIDIA, ASIC, HBM.
Source attribution: curated from Latent.Space (@latentspacepod). Cross-source corroboration: 10 independent sources have covered this development (including IT之家, Decoder, NVIDIA AI), confirming industry-wide signal strength.
