<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>KernelBench on 兜忆轩 | Cedric's Compiler Notes</title><link>https://douyixuan.github.io/tags/kernelbench/</link><description>Recent content in KernelBench on 兜忆轩 | Cedric's Compiler Notes</description><image><title>兜忆轩 | Cedric's Compiler Notes</title><url>https://douyixuan.github.io/images/papermod-cover.png</url><link>https://douyixuan.github.io/images/papermod-cover.png</link></image><generator>Hugo -- 0.146.0</generator><language>en-us</language><lastBuildDate>Thu, 08 Oct 2026 12:11:00 +0800</lastBuildDate><atom:link href="https://douyixuan.github.io/tags/kernelbench/index.xml" rel="self" type="application/rss+xml"/><item><title>023 · 从两百轮里挑最快的，算稳定胜过编译器吗？</title><link>https://douyixuan.github.io/posts/ai-silicon-papers/2026-10-08/</link><pubDate>Thu, 08 Oct 2026 12:11:00 +0800</pubDate><guid>https://douyixuan.github.io/posts/ai-silicon-papers/2026-10-08/</guid><description>CUDA Agent 在执行反馈环境中训练 CUDA 优化策略，并在 KernelBench 上显著超过 torch.compile；但评测允许每题最多 200 轮并选取轨迹中的最快正确解，策略能力与搜索预算仍需分开理解。</description></item><item><title>018 · Kernel 快了，端到端为什么几乎没动？</title><link>https://douyixuan.github.io/posts/ai-silicon-papers/2026-10-03/</link><pubDate>Sat, 03 Oct 2026 12:52:00 +0800</pubDate><guid>https://douyixuan.github.io/posts/ai-silicon-papers/2026-10-03/</guid><description>这篇论文把 kernel benchmark 的局部加速与真实 workload 的时间占比相乘：在 Transformer 中，可被生成式优化触及的部分仅占 8.9%–16.8%，因此看似不错的 kernel speedup 往往只能投射为约 1%–2.5% 的端到端收益。</description></item></channel></rss>