Linux kernel maintainers are preparing to remove support for AMD’s unreleased NPU2 in the upcoming Linux 7.0 release, a low-profile change that says a lot about how open-source software treats unfinished hardware. The move, proposed by AMD engineers themselves, reflects a preference for clean, proven code over features tied to products that never reached users.
It’s a small deletion with a bigger message.
An Unreleased Chip That Never Made It to Market
The hardware at the center of the change is AMD’s so-called second-generation neural processing unit, or NPU2.
It was never sold. Never shipped. Never announced in a consumer product. Yet its support code has lived quietly inside the Linux kernel for some time, waiting for hardware that ultimately didn’t arrive.
According to reports first highlighted by Phoronix, AMD engineers submitted patches to remove NPU2-related code from Linux 7.0. The reasoning was straightforward. Maintaining drivers for hardware that does not exist, and likely never will, adds overhead with no upside.
Kernel maintainers agreed.
This wasn’t a rejection or a controversy. It was more like a collective nod and a broom.
Why Linux Cares So Much About Dead Code
The Linux kernel is famously conservative about what it carries.
Every line of code adds maintenance cost. Every unused driver creates surface area for bugs, security issues, and confusion for future developers. That discipline becomes stricter with each major release.
Linux 7.0 represents a symbolic milestone, even if the numbering itself doesn’t imply a radical rewrite. Cleaning out abandoned hardware support fits naturally into that cycle.
Kernel developers often describe this philosophy bluntly. If no one can test it, no one can trust it.
NPU2 fit that category perfectly.
AMD’s Own Engineers Pulled the Plug
What makes this case notable is who initiated the removal.
The patch wasn’t pushed by an outside maintainer frustrated with vendor clutter. It came from AMD engineers, the same people who originally contributed the code. That detail matters.
It signals that AMD no longer expects NPU2 to surface in any future product. Keeping the driver around would only mislead developers and users into thinking the hardware might still appear.
By removing it themselves, AMD avoided dragging the kernel community along with outdated assumptions.
That kind of cleanup doesn’t always happen voluntarily.
What NPU2 Was Supposed to Be
NPU2 was intended as a follow-up to AMD’s first neural processing unit used in some Ryzen AI platforms.
Those NPUs are designed to handle on-device machine learning tasks more efficiently than CPUs or GPUs, such as voice processing, image enhancement, and lightweight inference workloads. The idea is similar to Apple’s Neural Engine or Qualcomm’s AI blocks.
The first-generation AMD NPU made it into limited shipping products.
NPU2 did not.
Sources familiar with AMD’s internal planning suggest the second-generation design ran into shifting priorities rather than a single technical failure. The AI landscape moved quickly, and AMD adjusted course.
Instead of pushing a half-baked accelerator, the company leaned harder into software optimization and broader heterogeneous computing strategies.
Sometimes the fastest move is stopping.
A Snapshot of the Industry’s AI Whiplash
The NPU2 episode mirrors a wider pattern across the semiconductor industry.
AI hardware plans change fast. Roadmaps are redrawn mid-cycle. Features that look essential one year become redundant the next as software frameworks evolve or customer demand shifts.
Vendors often experiment aggressively, knowing some designs won’t survive contact with the market.
Linux, however, doesn’t have the luxury of speculative optimism. Its job is to support what exists, not what might exist.
That tension explains why preemptive driver inclusion can backfire.
The Cost of Supporting Hardware That Doesn’t Exist
Even dormant code needs care.
Kernel APIs change. Subsystems evolve. Each new release risks breaking untested drivers in subtle ways. Maintaining NPU2 support would require ongoing updates, reviews, and regression checks, all for hardware no developer can plug in.
That effort pulls attention away from real users and real devices.
Kernel maintainers tend to ask a simple question. Who benefits?
In this case, the answer was nobody.
What This Means for AMD Going Forward
The removal of NPU2 support doesn’t signal a retreat from AI.
Quite the opposite.
AMD continues to invest heavily in AI acceleration through GPUs, CPUs with integrated AI features, and software stacks that lean on existing hardware. The company has also been vocal about supporting open standards and frameworks that don’t rely on niche blocks alone.
Dropping NPU2 suggests a focus on what can actually be delivered and supported at scale.
From a developer’s perspective, that clarity helps.
Lessons for Open-Source and Hardware Vendors
There’s a quiet lesson here for both sides.
For hardware vendors, upstreaming drivers early can be valuable, but only if the hardware roadmap is stable enough to justify it. Otherwise, those contributions risk becoming liabilities rather than assets.
For open-source projects, this episode reinforces the value of periodic pruning. Code removal is not failure. It’s maintenance.
Linux has survived decades precisely because it knows when to let go.
Linux 7.0 as a Line in the Sand
Major kernel releases often come with symbolic cleanups.
Linux 7.0 will include many changes more visible than the removal of an obscure AI accelerator driver. But moments like this show how the project stays healthy long-term.
No drama. No press release. Just a patch that deletes code nobody needs.
In open source, restraint is a feature.
A Ghost That Won’t Haunt Future Kernels
Once Linux 7.0 lands, NPU2 will quietly disappear from the kernel tree.
No users will notice. No systems will break. And that’s the point.
The “ghost hardware” won’t linger, confusing documentation or wasting developer time. Instead, the kernel moves forward with support focused on real, shipping devices.
In a space obsessed with the next big thing, sometimes the smartest move is acknowledging what isn’t coming.
