Linux 7.3-rc1 spends a third of its diff on AMD GPU register headers

August 30, 2026. Linus Torvalds closed the Linux 7.3 merge window by tagging 7.3-rc1. Two numbers frame the cycle: a release candidate of nearly 41 million lines of code, one of the largest openings in recent memory, and a diff in which a third comes from the AMD graphics driver alone. That is not a statistical fluke — it is a kernel swallowing next-generation hardware before it even gets to features.

The calendar, by contrast, is routine. 7.2 shipped on August 17, 2026, the merge window opened right after, and the 7.3 final release is due late October, most likely October 18 or 25. In between sit roughly seven weeks of release candidates, which is where the regressions flagged here are supposed to get fixed.

A third of the patch: the DCN 6.0 register dump

The single largest part of the diff is not executable code but register definitions: the DCN 6.0 (Display Core Next) headers for AMD’s upcoming display engine. The in-tree AMD graphics driver now runs to roughly 6.5 million lines, inflated by these hardware-constant files. Linus says it himself in the announcement: the register dumps, “as usual,” account for about a third of the whole rc1 patch.

This is neither waste nor a first. AMD replays the scenario every generation: it pushes hardware definitions upstream so its future cards are supported the day the kernel ships, without waiting another cycle. The trade-off is permanent background noise — the cycle’s real feature volume only becomes visible once you mask that block out, as Linus himself suggests. For anyone who reads LKML, a one-third “AMD GPU headers” share is now a routine marker, not an event.

Zen 6 and the early-enablement strategy

Beyond the display registers, 7.3 begins enabling the AMD Zen 6 platform. The goal is the same as on the GPU side: that the stable kernel due in late October 2026 already knows how to drive silicon arriving a few months later. For testers on current hardware, that is a useful window; for everyone else, it is a reason not to mistake diff size for immediate value.

Linux gamers gain more than anyone

For end users, the most concrete changes in 7.3 land on the gaming side. Three stand out.

A native driver for the Steam Controller. The HID subsystem merges patches from Vicki Pfau, a Valve contributor, giving the original Steam Controller a native in-kernel driver. Until now, using it without the Steam client required SDL2 or third-party workarounds; now it works as a standard input device, whether Steam is running or not. A modest win, but it removes fifteen years of friction for players who never install the client.

Better vRAM overcommit handling. The DMEMCG work led by Natalie Vock, a Valve engineer, improves how the AMDGPU driver evicts video memory when a game asks for more than you have. In the cited testing, a game pushed toward 9 GB of vRAM on an 8 GB card stayed playable — about 1 GB of overcommit absorbed without a crash. The mechanism hits a real pain point: 8 GB cards, now the entry-level norm, against titles that demand more and more memory. The work is described as ongoing, so further gains are expected after 7.3.

A scheduling fix for gaming. The “flatten the pick” patch series by Peter Zijlstra (Intel) attacks cgroup scheduling: it fixes a case where accumulated background weight starved the foreground process of CPU slices. Benchmarked on a Sandy Bridge paired with a Radeon RX 580, it yields a measurable FPS gain and lower frametime variance — exactly the scenario of a game competing with a loaded browser or a Discord call. It joins a load-balancing fix for Intel’s hybrid Core Ultra CPUs, a regression that had lingered for years.

The downside: a Nouveau regression

The cycle is not all good news. Early 7.3 benchmarks show graphics regressions on Nouveau, the open-source Nvidia driver. Caught early, they have time to be fixed before the stable release — but if you game on Nvidia hardware using the free driver rather than the proprietary one, that is reason enough to stay off the 7.3 release candidates.

The rest: KVM, Apple, Btrfs, and quiet foundations

The cycle also carries a rework of KVM guest memory conversion, initial USB4 and Thunderbolt support for Apple M1 through M3 chips (with a new PMGR power-management driver that handles gating of on-chip blocks), out-of-the-box Intel Xe3P (Nova Lake) graphics, Vulkan video decode for NVK, Btrfs performance gains from David Sterba, and FailFS, a minimal filesystem that always returns an error — useful for exercising error-handling paths. Underneath, the cycle also cleans house: the ancient efs and freevxfs filesystems are removed from the tree. Nothing flashy, but a foundation that sets up the October release.

Why the AMD driver dominates the tree

The 6.5 million lines of AMD graphics driver is not just a statistical curiosity — it is the direct consequence of a strategic choice made years ago. AMD bet on a fully open-source amdgpu driver maintained in-tree, where Nvidia keeps the bulk of its driver proprietary and exposes only a fraction. The result: all of AMD’s hardware support — registers, power management, encoding — lands in the kernel itself, mechanically inflating the diff with every generation.

That choice costs readability but pays a concrete dividend to users: a Radeon card works out of the box on any distribution, with no module to compile and no driver to install. It is also why Valve invests in amdgpu for the Steam Deck and why the vRAM gains in 7.3 directly benefit the handheld. The diff inflation is the price of that strategy, and it will only grow as long as AMD keeps pushing silicon upstream.

How to test the candidate

If you do want to run 7.3-rc1, the practical route is a rolling distribution or a mainline build: grab the tarball from kernel.org, or pull the linux tree from git.kernel.org and build with your existing config. Arch, openSUSE Tumbleweed, and the kernel-testing spins all track mainline closely. Keep a known-good kernel available as a fallback, since an rc1 is by definition unproven — the Nouveau regression is a reminder that early benchmarks catch things a release notes diff never will.

Verdict

If you run AMD or recent hardware, the 7.3-rc1 is worth testing for early Zen 6 enablement and the vRAM gains. If you game on Nvidia with Nouveau, stay off the candidates until the regression is confirmed fixed. If you are an ordinary user, wait for the late October stable release — that is where the regression fixes and feature stabilization will have done their work.

References

• Linux 7.3-rc1 — Linus Torvalds’ announcement, August 30, 2026
• Linux 7.3-rc1 With AMD Zen 6 Additions, Better Btrfs Performance & 2026 Steam Controller — Phoronix, August 30, 2026
• What Linux Kernel 7.3 Actually Changes for Gamers — Digitnaut, August 30, 2026
• Linux 7.3-rc1 Lands With AMD Zen 6, a Native Steam Controller Driver, and Apple Thunderbolt — Wimantis
• Linus Torvalds Announces First Linux Kernel 7.3 Release Candidate — 9to5Linux