# Linux 7.5 boots with AMDXDNA NPU, Mali G68 and HDMI 2.0

> The first Linux 7.5 pull requests are already in DRM-Next: the Ryzen AI NPU now reports its usage, Panfrost adds Mali G68 and Rockchip gains HDMI 2.0.
> By José Maldonado · Technology · 2026-11-04

The next kernel version already has work in the queue while the previous one is still in the workshop. The first graphics driver pull requests for Linux 7.5 landed in DRM-Next on October 29, with the 7.4 merge window still open.

The content is not a roadmap: these are patches sent to the tree where graphics drivers are cooked before reaching the kernel. The submissions touch the AI accelerator AMD sells under the Ryzen AI brand, bring the Arm Mali G68 GPU into the open-source Panfrost driver and bring HDMI 2.0 to Rockchip chips. The direction is the same in all three cases: hardware that depended on out-of-tree software is starting to have its own path in Linux.

The calendar helps gauge the size of what is coming. Linux 7.4-rc1 came out on Sunday, November 1, with nearly twenty thousand commits, well above the usual thirteen or fourteen thousand, and Linus Torvalds described it as a huge merge window. The 7.5 changes slip into an already loaded cycle, and this time the main novelty is not more raw power, but something more modest and considerably more useful: being able to see what the NPU is doing while it works.

## A cycle that starts ahead of time

DRM-Next is the integration tree where Dave Airlie, head of the graphics subsystem, collects each cycle's pull requests from maintainers before sending them to Linus Torvalds. What travels now is not a release, but patches waiting in that queue.

That 7.5 material arrives now is not an anomaly: the drm-misc-next team prepares its branch while the previous window closes, so the first submission goes in as soon as the next one opens. The first batch of this round appeared days after the 7.4 window opened and the second arrived the following week.

The numbers from the previous cycle provide the context. Between Linux 7.3 and Linux 7.4-rc1, 19,671 commits accumulated, the highest figure recorded in a merge window, compared with 16,267 in the previous cycle according to Phoronix's count. It is the same cycle we already covered when discussing [the largest merge window in its history](https://observatorioblockchain.com/en/technology/linux-7-4-rc1-ventana-merge-mas-grande/), and Torvalds attributed much of the volume to work generated by AI tools.

None of this is in a stable kernel. These are patches in the integration queue, and the distance to the final version is measured in weeks of release candidates and in reviews that could still reject any of the changes.

## AMD's NPU stops being a black box

An NPU is a processor specialized in neural network operations, designed for low-power inference, distinct from the CPU and GPU although it shares memory with them. On Linux, AMD's NPU is handled by the AMDXDNA driver, which entered the kernel with version 6.14 in early 2025, according to [the pull request that introduced it](https://lore.kernel.org/lkml/CAPM=9tw+ySbm80B=zHVhodMFoS_fqNw_v4yVURCv3cc9ukvYYg@mail.gmail.com/).

What is coming in 7.5 seems lesser: the driver now reports the time the NPU is busy. Max Zhen's patch, sent on October 12 to the kernel mailing lists, accumulates the nanoseconds of execution for each process and publishes them in fdinfo, the /proc file that describes an open descriptor, under a key of the drm-engine type. With that, any tool that reads that file can calculate the occupancy percentage without relying on driver-specific ioctls.

The fine print matters more than it seems. On October 23, the same engineer sent a correction so that the key would match the one already used by the driver's out-of-tree applications, because several of them read it by name. A compatibility detail like that decides whether a monitoring tool shows data or an empty line, and explains why observability work does not end when the counter works.

Measuring should not be confused with speeding up. The patch does not make inference faster or improve the performance of a specific model: it places a counter where there was previously silence, and that serves to diagnose bottlenecks, not to promise figures.

## AIE4, the generation that has not yet reached store shelves

The second submission is more substantial. The driver adds AIE4 features, the next generation of AMD's NPU block: reading and adjusting the power mode, managing the lifecycle of kernel queues, configuring the memory layout and suspend with resume. The work goes back a long way, because since March 2026 there have been patches from AMD engineers for AIE4 on the public mailing lists.

The question that still has no answer is which processor it will debut with. Phoronix describes AIE4 as the NPU of the next generation of Ryzen AI and points to Zen 6 platforms, with the caution of someone who has no confirmation from AMD itself. It is an inference from the calendar, not an announcement, and it should be read as such before planning a purchase.

Suspend and resume sounds like an engineering detail, but it is the difference between an NPU that remains available when the laptop lid is closed and one that disappears until the next boot.

## Mali G68, what happens when the kernel looks at mobile devices

Panfrost is the kernel driver for Arm Mali GPUs and the free alternative to Arm's driver that never made it into the mainline tree. Mesa provides the user-space part. With the G68, Panfrost adds a GPU from the Valhall family that appears in Samsung's Exynos 1280 and in several MediaTek Dimensity chips, including the 900 and the 1080.

Therein lies the detail that goes beyond the desktop. The Mali G68 is not a graphics card GPU, but a component integrated into mid-range phones, tablets and small boards, such as the Galaxy A33 and A53 5G or the Redmi Note 12 Pro, according to the spec sheet from [Notebookcheck](https://www.notebookcheck.net/ARM-Mali-G68-MP4-GPU-Benchmarks-and-Specs.566225.0.html)They are precisely the devices that stop receiving manufacturer support after three or four years and for which a mainline kernel with Panfrost is the way to keep receiving patches.

The submission also includes improvements in reset handling and driver initialization. On a GPU that hangs, the difference between recovering and ending up with a black screen is exactly that, and on devices without a service console a clean reset is the only safety net left.

## HDMI 2.0, pixel format and a new panel

The rest of the package has fewer headlines and more users. Rockchip's dw-hdmi-qp driver gains HDMI 2.0, which opens the door to 4K at 60 Hz on boards with the company's SoC; VirtIO GPU adds the ARGB8888 format on primary planes, useful for those working with virtual machines; and the BOE NE140WUM-N41 panel arrives for laptops that until now booted with the screen in basic mode.

Change

Driver

Who it affects

NPU occupancy time

AMDXDNA

Those measuring local inference on Ryzen AI

Power modes, queues and AIE4 suspend

AMDXDNA

Laptops with AMD's next NPU

Mali G68 and cleaner reboots

Panfrost

Mobile phones, tablets and Arm boards

HDMI 2.0

dw-hdmi-qp (Rockchip)

Boards with Rockchip SoCs

ARGB8888 on primary planes

VirtIO GPU

Guests with accelerated graphics

The table leaves a clear pattern: almost everything coming in is small enablement work that shows up on a specific machine. That is how a kernel cycle works when the hardware is already on the street and the software follows behind, with patches that fill gaps rather than redraw the map.

## What to watch when the 7.5 window opens

The calendar places the 7.5 merge window at the end of the year, and until then there are two fronts. The first is the stabilization of 7.4, where the graphics and storage changes that have already landed, such as [the large ext4 writes with iomap](https://observatorioblockchain.com/en/technology/ext4-iomap-linux-7-4-escrituras/)they need to get through the weekly candidates without any scares. The second is whether the AIE4 work arrives intact or stalls at the door, something that has already happened with hardware that does not exist in stores.

For the user, the value of this cycle is less impressive than it sounds. An NPU with a usage counter and a Mali G68 recognized by the kernel do not move a benchmark, but they broaden the list of devices that work with free software without depending on a manufacturer. Those who game on Linux have already seen how [the scheduler and VRAM management in 7.3](https://observatorioblockchain.com/en/technology/linux-7-3-failfs-scheduler-gaming-vram/) changed perceived performance: here the move is similar, except that it is measured in compatibility and in years of support.
