Together AI 发布 ThunderKittens：简化 AI 内核编程的嵌入式 DSL

AI 摘要

Together AI 发布了名为 ThunderKittens 的开源嵌入式 DSL，旨在简化 AI 内核的 CUDA 编程，使 GPU 性能最大化。该框架在 H100 上比 FlashAttention-2 更快，并已集成到 NanoGPT 项目中。团队希望降低 AI 内核开发门槛，同时保持对底层硬件的透明控制。

正文节选

We have a lot of fun building kernels for AI, and we’re proud of our open-source AI contributions. We’re less proud that people regard FlashAttention as alien technology. In a NeurIPS keynote, our group tried our best to make the key ideas simple–but the gap from the beautiful pictures generated by Dan Fu to actual “gpu go brr” cuda is still too damn high. So we set out to write a simple framework to make it really easy to express the key technical ideas. Relatively quickly, we had a small libra

发布时间：—

抓取时间：2026-08-02 00:28

来源机构：Together AI