Linux Kernel 7.5 drops the last 30-year-old Intel MMX RAID code, replacing 800+ lines with 12 lines of scalar C. See what changes for 32-bit CPUs.
-
Published by Kunal Chowdhury
on 18 November, 2026
Add kunal-chowdhury.com on Google to see more of our coverages in your search results.
If you grew up in the late 1990s, you probably remember the holographic "Intel Inside" stickers promising multimedia magic on shiny beige Pentium MMX towers. Nearly 3 decades later, the mainline Linux kernel is finally pulling the plug on the last piece of code written specifically for those chips.
Staged in Andrew Morton's kernel tree for Linux 7.5, a new 3-patch series from Google engineer Eric Biggers retires the kernel's final 32-bit Intel MMX optimizations. By purging legacy software RAID XOR routines, Linux replaces over 800 lines of 1990s assembly-era code with just 12 lines—marking the quiet end of a 30-year hardware chapter.
memcpy(), copy_page(), and clear_page() back in 2021, leaving RAID as the lone holdout.
Intel unveiled MMX in January 1997 on the P5-based Pentium MMX, quickly following up with the Pentium II in May 1997 and rival implementations like the AMD K6. Marketed to consumers as "MultiMedia eXtensions" for accelerating CD-ROM video, audio, and 2D graphics, MMX gave x86 its earliest mainstream taste of Single Instruction, Multiple Data (SIMD) parallelism—even though court filings during a trademark dispute with AMD later revealed "Matrix Math Extensions" as another internal meaning.
By the time Linux 7.5 ships in 2027, MMX will hit its 30th birthday. Back in the late 1990s, doubling integer throughput on packed 64-bit chunks felt transformative for software decoders and early Linux multimedia tools. Today, modern desktops run Ubuntu 26.10 with amd64-v3 AVX2 optimizations or Linux Lite 8.2 on Kernel 7.3, turning hand-tuned 1997 assembly inside the mainline kernel into pure technical debt.
As spotted by Phoronix's report on the Linux 7.5 MMX removal patches, almost every other trace of kernel MMX disappeared half a decade ago. Back in 2021, commit c6dbd3e5e69c (x86/mmx_32: Remove X86_USE_3DNOW) already purged MMX optimizations from heavily called memory routines like memcpy(), copy_page(), and clear_page().
That left just a single outlier in the entire tree: 32-bit software RAID XOR acceleration, used exclusively on 25-to-30-year-old CPUs that have MMX but lack SSE support—and only when someone builds a RAID array on them. According to Tom's Hardware's breakdown of the cleanup, Biggers' 3-patch series wipes out more than 800 lines of MMX RAID code, replaces them with 12 lines, and cleans up legacy CPU feature checks in RAID6.
Anyone maintaining a 32-bit kernel subsystem knows how fragile dead-code branches become when almost nobody tests them on real silicon. Removing those unmaintained assembly blocks shrinks the attack surface, eliminates obsolete conditional macros, and keeps the RAID6 parity dispatcher clean for modern architectures.
Beyond shrinking maintenance overhead, kernel engineers have a deep architectural reason to celebrate seeing MMX go. To avoid forcing 1990s operating systems to rewrite their task-switching code, Intel aliased MMX's 8 64-bit registers (MM0 through MM7) directly onto the mantissa bits of the 80-bit x87 floating-point (FPU) stack.
Because of that clever hack, a processor could not execute x87 floating-point math and MMX integer instructions at the same time without risking register corruption and requiring an EMMS reset. Inside kernel space—where FPU state saving is strictly guarded—supporting MMX fast paths added disproportionate complexity for 0 cryptography or CRC benefits.
When Intel introduced SSE in 1999 with the Pentium III, it finally gave CPUs independent 128-bit XMM vector registers that never collided with the x87 stack. Every modern Linux SIMD routine since then has targeted SSE2, AVX2, or AVX-512 instead, leaving 32-bit MMX stranded in a bygone era.
Looking at how far x86 vector processing has evolved between 1997 and the upcoming Linux 7.5 release puts this retirement in sharp perspective:
If you keep a vintage Pentium II box humming in a closet with a modern kernel, do not panic. Removing MMX acceleration from the RAID XOR driver does not remove software RAID support itself—the kernel simply routes parity calculations through its clean, well-tested scalar C fallback instead.
Nor does this change affect user-space software. Any 32-bit retro binary, emulator, or desktop tool running on an upstream GNOME 51 Linux setup or an older 32-bit system can still execute MMX instructions in user mode if the silicon supports them. Still, alongside Linux 7.0 dropping the 28-year-old Intel 440BX EDAC driver and recent 32-bit patches making the Time Stamp Counter (TSC) mandatory, the message is unmistakable: Linux is streamlining its core for the next 30 years.
At the same time, Linux 7.5 is pouring fresh engineering effort into modern silicon. Alongside retiring 1997 MMX routines, the upcoming kernel brings ChaCha AVX-512 optimizations that deliver up to 99% faster encryption throughput on AMD Zen 5 processors, plus expanded AMD Ryzen AI NPU enablement.
Retiring the last 800 lines of 1997 Pentium MMX code may barely register on a modern benchmark chart, yet it feels like a genuine passing of the torch for anyone who built PCs in the 1990s. By clearing away 30-year-old x87-aliased assembly paths, Linux 7.5 becomes leaner, safer to maintain, and sharper for modern AVX-512 hardware. What was the very first MMX-branded processor you ever owned or wrote code for? Share your nostalgia in the comments below!
Solution Architect & Former Microsoft MVP
Kunal Chowdhury is an enterprise solution architect and former multi-year Microsoft MVP. He is the author of three technical books: Windows Presentation Foundation Development Cookbook, Mastering Visual Studio 2017, and the Mastering Visual Studio 2019.
He publishes technical articles on Kunal-Chowdhury.com.