跳到主要內容

Claude Code評測:Torvalds用AI改寫Linux核心效能 | Claude Code Review: Torvalds Uses AI on Linux Kernel

By Kit 小克 | AI Tool Observer | 2026-10-13

🇹🇼 Claude Code評測:Torvalds用AI改寫Linux核心效能

Claude Code這次真的碰到Linux核心了

Claude Code 最近立了一個不算小的戰功:Linux 創始人 Linus Torvalds 親自用它改寫核心最底層的檔案讀取路徑,而且效能實測提升將近三倍。這不是哪家公司的行銷稿,是10月12日直接貼上 Linux kernel mailing list(LKML)的真實 patch 系列,任何人都能去看原始 commit。

發生了什麼事

Torvalds 在自己的 git 分支 noref-filemap-read-with-llm 上推了一組 8 個 patch,標題是「mm: read from the page cache without folio references」。他在 commit 訊息裡老實交代:程式碼、測試程式、code review,甚至 cover letter 的初稿都是 Claude Code 寫的,他自己負責下指令、盯方向,並在自己的筆電上跑 benchmark 驗證數字。

數字沒有灌水

這次要解決的問題是:當多個執行緒同時小量讀取同一個檔案時,核心原本得對每個 folio(分頁快取單位)做 atomic reference count,造成嚴重的 cache line 競爭。新作法改用 RCU read-side critical section 跳過這個計數步驟,實測結果:

  • 小量讀取的每秒讀取數,從約 990 萬~1040 萬次,衝到 3000 萬~3300 萬次,接近三倍
  • cache line 競爭事件,從 30,538 次暴跌到 101 次

誠實講,先別太興奮

這條分支目前還沒有合併進 Linux 主線,疊在 7.3-rc6 上面,Phoronix 認為有機會趕上 7.4 合併窗口,但目前只在 Torvalds 自己一台筆電上測過,樣本數等於 1,離正式進核心還有審查與其他硬體驗證要跑。Torvalds 自己也強調,是他在「導演」整個過程——AI 不是自主決定要改哪裡,也不是自己判斷對不對,他才是最後把關的人。

這件事真正的意義

比起「AI 會寫 Linux 核心了」這種標題殺手,更值得注意的是:連全世界對程式碼最挑剔、最不買行銷話術的工程師之一,都願意公開承認自己拿 Claude Code 當助手。對還在猶豫要不要把 AI coding agent 放進正式開發流程的團隊來說,這是一個很現實的參考點,而不是又一個廠商的宣傳案例。

好不好用,試了才知道。


🇺🇸 Claude Code Review: Torvalds Uses AI on Linux Kernel

Claude Code Just Touched the Linux Kernel

Claude Code scored a notable win this week: Linux creator Linus Torvalds used it to rewrite one of the kernel's lowest-level file-reading paths, and the benchmarked result was almost a 3x throughput gain. This isn't a vendor press release — it's a real patch series posted to the Linux kernel mailing list (LKML) on October 12, and anyone can read the raw commits.

What Actually Happened

Torvalds pushed an 8-patch series titled "mm: read from the page cache without folio references" to his own noref-filemap-read-with-llm git branch. In the commit messages, he's upfront about it: the code, test programs, code review, and even the first draft of the cover letter were written by Claude Code, while he directed the work and ran the benchmarks himself, on his own laptop.

The Numbers Aren't Hype

The problem being solved: when many threads do small reads from the same file, the kernel has to bump an atomic reference count on each folio (the page cache unit), creating heavy cache-line contention. The fix skips that counting step using an RCU read-side critical section instead. Measured results:

  • Small-read throughput jumped from roughly 9.9–10.4 million reads/sec to 30–33 million reads/sec — close to 3x
  • Cache-line contention events dropped from 30,538 to 101

The Honest Caveats

This branch is not merged into mainline Linux yet — it sits on top of 7.3-rc6. Phoronix suggests it could land in the 7.4 merge window, but so far it's only been benchmarked on Torvalds' own single laptop — a sample size of one, with more review and hardware testing still ahead. Torvalds himself is clear that he was directing every step: the AI didn't decide what to change or judge whether the result was correct — that part was still entirely on him.

Why This Actually Matters

Skip the "AI can write the Linux kernel now" headlines. The real signal is that one of the most notoriously skeptical, marketing-allergic engineers on the planet is willing to publicly credit Claude Code as his assistant on real systems code. For teams still on the fence about putting AI coding agents into their actual workflow, that's a far more grounded data point than another vendor demo.

You won't know until you try it.

Sources / 資料來源

延伸閱讀 / Related Articles


AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

留言

這個網誌中的熱門文章

Google Ironwood TPU v7 推理專用晶片解析:效能追平 NVIDIA、成本低 44%,AI 晶片戰爭正式開打 | Google Ironwood TPU v7 Explained: Matching NVIDIA Performance at 44% Lower Cost — The AI Chip War Heats Up

Claude Code 實測:AI 幫你寫程式到底行不行? | Claude Code Review: Can AI Really Code for You?

Cursor vs GitHub Copilot vs Claude Code:AI 程式助手大比拼 | AI Coding Assistants Compared: Cursor vs GitHub Copilot vs Claude Code