--- title: "Linux 7.5: kernel ChaCha AVX-512 code up to 99% faster on AMD Zen 5" description: "Linux 7.5 is set to rewrite ChaCha's AVX-512 code, which was tuned for early Intel Skylake chips, and Phoronix reports gains of up to 99% on AMD Zen 5. The kernel's cipher was still tuned for chips that slowed down the moment you used AVX-512." canonical: https://inyourgeek.com/en/articles/2026-11-09-linux-7-5-le-chacha-avx-512-du-noyau type: article author: "Sébastien Soulier" publisher: "InYourGeek" language: en datePublished: 2026-11-09 dateModified: 2026-11-09 section: "Software" tags: ["linux", "chiffrement", "avx512", "amd"] --- # Linux 7.5: kernel ChaCha AVX-512 code up to 99% faster on AMD Zen 5 > Linux 7.5 is set to rewrite ChaCha's AVX-512 code, which was tuned for early Intel Skylake chips, and Phoronix reports gains of up to 99% on AMD Zen 5. The kernel's cipher was still tuned for chips that slowed down the moment you used AVX-512. November 9, 2026 · Software · 3 min ## Code tuned for a chip that ran out of breath [Phoronix](https://www.phoronix.com/news/ChaCha-AVX-512-Linux-7.5) reported it on Monday 9 November 2026: Linux 7.5 will retune the code for ChaCha, the Linux kernel's symmetric stream cipher. That code has long had an optimised path for AVX-512, the wide vector instructions on x86 processors. But that path was built for the first Intel generation to offer those instructions: Skylake. Those Skylake chips had an awkward habit. Their clock speed dropped as soon as you put them to work on AVX-512. So the kernel code was tuned for those chips, with that limitation in mind. It was a sensible call at the time. Then processors moved on, and the code stayed put. That's the usual fate of an optimisation: you write it for the machine in front of you, and it outlives that machine, like a sticky note still clinging to a monitor that was recycled years ago. ## A retune for today's chips The change planned for Linux 7.5 reworks that code so it runs better on current Intel and AMD processors with AVX-512, which no longer suffer from that frequency drop. Put simply, the kernel has been tiptoeing around a sleeping roommate who moved out ages ago, and it's finally allowed to turn the music up. The headline figure is spectacular: on current-generation AMD Zen 5 processors, ChaCha becomes up to 99% faster. In the best case, its throughput nearly doubles. Mind the words *up to*: that's a ceiling, not an average. It's the number printed on the box, not the one you get on a wet Tuesday afternoon. The source gives no comparable figure for recent Intel processors. ## What it changes, and what the source doesn't say For kernel-side encryption, the direct effect is simple: on a recent machine with AVX-512, and especially on Zen 5, ChaCha will encrypt and decrypt noticeably faster. Only the implementation changes. The algorithm stays the same. ChaCha is still ChaCha, the same dance, except the kernel has stopped asking the processor to perform it in slow motion. The grey areas matter as much as the headline number. The source doesn't say which kernel workloads will actually benefit. It doesn't give the average gain either, or how results vary across processor generations. Nor does it say what happens to the early Skylake chips, the very ones the old tuning was written for. Finally, nearly doubling encryption throughput won't make your whole system twice as fast: encryption is just one step among many. Before you promise anything to your CIO, wait for benchmarks on your own workloads. “Phoronix said 99%” is not a capacity plan. ## Technical debt, silicon edition This episode is a useful reminder for anyone who optimises low-level code. An optimisation is never neutral. It rests on assumptions about the hardware, and those assumptions age without telling anyone. The code keeps working, the tests pass and nothing breaks. It just runs slower than the machine allows, and nobody notices until someone measures. It's probably the stealthiest kind of technical debt. No bug report gets filed and no error shows up in the logs. The only clue is the gap between what the code does and what the processor could do, and here that gap reaches, at best, 99%. The kernel spent years going easy on a processor that got winded every time it used its own vector instructions. Turns out the chips hit the gym a while ago, and the software was the last to hear about it. ## Sources - [ChaCha AVX-512 Optimization For Linux 7.5 Yields Up To 99% Improvement On AMD Zen 5](https://www.phoronix.com/news/ChaCha-AVX-512-Linux-7.5) — phoronix.com _Article written with AI assistance from the cited sources, reviewed and approved by a human before publication._ --- Equivalent HTML page : https://inyourgeek.com/en/articles/2026-11-09-linux-7-5-le-chacha-avx-512-du-noyau Published by InYourGeek — https://inyourgeek.com/en/a-propos Method and safeguards : https://inyourgeek.com/en/methode All articles : https://inyourgeek.com/en/archives