# AMD brings RDNA 5 to Linux with 12,643 lines of GFX13

> AMD publishes eight patches and 12,643 lines for GFX13, the basis of RDNA 5, in the Linux kernel. What it already enables and what remains unconfirmed.
> By José Maldonado · Technology · 2026-11-11

A graphics card that does not exist yet has just received twelve thousand lines of code in the Linux kernel. On November 9, AMDGPU maintainer Alex Deucher sent a series of eight patches to the kernel mailing list to enable GFX13, the architecture everyone assumes will be RDNA 5. They are 12,643 lines spread across 32 files, with 15 deletions, and they do not include the register headers that usually accompany this kind of work. No one outside AMD has seen the silicon that this code will handle.

The submission is no surprise. GFX13 support has been in the LLVM compiler since January, Display Core Next 6 arrived in August and the GDDR7 patches, the memory expected in the next generation, appeared in September. What was missing was the central piece: the driver that makes the graphics engine work under Linux. That is what has now arrived, and it is also the clearest sign that the program is moving forward in the usual order, software first and silicon afterwards.

The useful reading for the Linux user is not about performance, because there is no performance to measure. It is about the calendar: when distributions receive the kernel that includes these patches, the support will exist before the card goes on sale. That advantage is what has allowed Linux to play with an edge in every GPU generation, and the reason Valve can move SteamOS so nimbly on hardware that has not yet launched.

## What exactly does the patch series enable?

The submission is not limited to registering hardware identifiers, which is what the first patches for a new architecture usually do. It includes the GFX13 graphics engine driver, new implementations of the Micro Engine Scheduler and the Integrated Management Unit, components of the Kernel Fusion Driver that manages compute workloads, a trap handler to save and restore compute threads and several clock control mechanisms. In other words, it lays the foundations for scheduling, compute and power management.

The [GFX13 patches sent by Deucher](https://www.phoronix.com/news/AMDGPU-Patches-GFX13) cover two specific targets, GFX 13.0.1 and GFX 13.0.4. Having two identifiers suggests at least two different hardware configurations in preparation, although it does not imply two commercial cards: they could be silicon variants under the same market name. One of the patch notes also acknowledges an explicit limitation, that MES queue reset is not yet supported and must remain disabled until it is developed.

That detail best describes where the work stands. It is not a finished driver, it is the first layer of an enablement that will be completed with later submissions over the coming months. Anyone who installs a kernel with these patches will not get a new GPU working, because there is no GPU to install.

The acronyms are worth translating, because the patch text is dry. The Micro Engine Scheduler is the scheduler that distributes work among the GPU queues; the Integrated Management Unit manages the powering on and off of internal blocks; the Kernel Fusion Driver is the piece that exposes the card as a compute device for calculation tasks. That all three appear in the first submission indicates that we are talking about a complete architecture.

A GPU driver is built in layers, and this submission is the first brick of the layer that matters. What comes now are fixes, testing on engineering silicon and, later, commercial names.

## The calendar that can be read between the lines

Adding up the pieces reveals a fairly coherent sequence. January 2026 leaves the GFX13 backend in LLVM ready to compile. August brings the DCN6 display engine. September adds the GDDR7 patches, which point to the type of memory for the next generation. November delivers the graphics engine. Each milestone arrives a few months apart, which is the pace at which AMD usually prepares an architecture before the first product appears.

The next foreseeable stop is kernel 7.5, although inclusion is not confirmed. If the patches enter that cycle, distributions that follow the stable kernel will have them available during the first half of 2027. The current cycle is at [Linux 7.4 rc2, which arrived with FAT32 up to 162 times faster](https://observatorioblockchain.com/en/technology/linux-7-4-rc2-fat32-162x-torvalds-ia/)so the window fits with the usual schedule. Rumors place the first RDNA 5 product in the second half of 2027, starting with a mid-range GPU called AT2, with the rest of the family sliding into 2028.

Milestones are not a coincidence; they are a method. Each intellectual property block is enabled in the kernel tree when it is ready, and the set is complete months before the card reaches a store.

For the Linux user, that means something concrete and not very glamorous: support will be there before it is needed. The practical consequence is that buying a card will not mean waiting six months for green drivers or putting up with a desktop that flickers while the driver matures. That quiet advantage depends on two separate pieces moving forward at the same time: the kernel driver and the work in Mesa, the library where the graphics API that games consume is implemented.

## What the patches do not say

There is no compute unit count, no clock speeds, no memory configuration, no performance figures. Nor are there any trade names. AMD has not officially confirmed that GFX13 will ship as RDNA 5, nor has it announced a launch date. All that can be stated with the patches in hand is that the company is preparing at least two graphics IP configurations and that the software work is running ahead of the hardware, as is usual.

That absence of data fuels speculation, and here it is worth being explicit: any performance figure circulating these days is invented or leaked without verification. The size of the patch, twelve thousand lines, says nothing about performance. It says the architecture has enough changes to justify a new driver, which was to be expected in a generational leap.

There is also a market reading that deserves qualification. Patch submissions usually precede silicon sampling by several quarters, so November 2026 fits with a commercial launch in 2027. AMD publishing the work openly allows distributions and independent developers to test and report before the product exists, something vendors that rely on proprietary blobs cannot offer.

## The driver before the silicon

The way AMD builds its drivers explains much of its position in Linux. Each generation lands in the main kernel tree, is reviewed in public and is tested on hardware nobody has yet. The same open approach brought [ROCm to FreeBSD](https://observatorioblockchain.com/en/technology/freebsd-integra-software-rocm-amd-hpc-ia/) for compute and AI workloads, and it is what allows a smaller distribution to have support on day one. The result is an ecosystem where the card works the day it is plugged in, without depending on an installer that only supports one distribution and one specific version.

It will be months before we know whether RDNA 5 delivers what the roadmap promises, and data is still lacking for any serious comparison with the competition. The only thing verifiable today is that the groundwork is done and that it can be audited line by line on the kernel mailing list, which is more than those who announce architectures with slides offer.
