The first release candidate of Linux kernel 7.3, Linux 7.3-rc1, was published on August 30, 2026. During the two-week merge window, Btrfs performance fixes, changes to CPU and GPU memory allocation, support for next-generation SoCs, and removal of old code all landed at once. But this is not a stable release. Rather than counting a mass of changes as a single "speedup," it's more accurate to separate out which users get what benefit, and when it actually kicks in.
Strip out a third of the patches, and the scale returns to normal
By commit count, 7.3-rc1 ranks among the largest in the kernel's history. But it isn't the largest ever. Roughly a third of the patches in Linux 7.3-rc1 consist of AMD DCN6 register headers and next-generation-related code, and Linus Torvalds explained that if you set that portion aside, the merge window was not much different from usual.
Register headers lay out numerous hardware control parameters as C-language definitions. They're necessary groundwork for running a new GPU generation, but the number of added lines doesn't correspond to the number of user-facing features. A huge volume of changes reflects maintenance burden, but it doesn't mean the kernel as a whole became proportionally more complex, or proportionally faster.
Moreover, 7.3-rc1 marks the boundary where feature merging ends and testing begins. From here, bugs get fixed and regressions get flushed out. For production environments, the decision to upgrade from Linux 7.2 shouldn't be based on the length of the feature list, but on verifying the stable release against your own workload and hardware first.
Btrfs's big improvements come with clearly defined applicable paths
The Btrfs changes come with both impressive numbers and clearly stated applicability conditions. Previously, direct I/O that failed to meet alignment requirements would fall back to buffered I/O, preserving correctness at the cost of performance. In 7.3, an iomap bounce buffer is used instead, processing the I/O without falling back to the page cache. According to the maintainer's pull request, the proportion of paths covered rose from a theoretical maximum of about 50% to about 95%.
Other reported sample workloads include roughly a 3x improvement from switching extent buffer tracking—used to throttle usage—from XArray to a local LRU list, and roughly a 5x improvement in fsync for files with many extents and no holes, under a non-SSD mode running multiple log operations simultaneously. That fsync operation is measured in microseconds. In each case, the gains come from targeting specific slow internal paths.
Real applications show mixed results. When Phoronix compared Linux 7.2 against the 7.3 development snapshot as of August 23, using a Xeon 678X, a 2TB PCIe 4.0 SSD, and default mount settings, Btrfs jumped from last place to first for TigerBeetle, a financial transaction database. MariaDB, on the other hand, showed a regression: with copy-on-write enabled under default settings, overall performance generally lagged behind EXT4, F2FS, and XFS. The roughly 5x figure is not a universal verdict for choosing a filesystem.
This discrepancy isn't a contradiction. The kernel changes eliminate specific latencies and lock contention, but applications have their own distinct I/O patterns, sync frequencies, and data layouts. What Btrfs users need to check isn't the average improvement rate, but whether their own workload—virtual machines, databases, multiple clients, and so on—actually passes through the improved path.
The scheduler and VRAM management change who gets the resources
On the CPU side, task selection across cgroup hierarchies has been restructured. The previous hierarchical selection method could let intermediate levels of the hierarchy obscure runnable tasks, causing problems with how tasks whose nice values had been changed were handled, and with latency. 7.3 consolidates tasks into a single run queue and dynamically reflects the hierarchy's weighting. The default weight-distribution scheme, concur, is the most precise, and the maintainer notes that this precision comes with a proportionally higher computational cost.
The same merge also includes changes that reduce latency for tasks running in short time slices, and fixes to load balancing on systems with CPU cores of differing performance. This is why gains are expected for gaming performance on older PCs and for newer hybrid CPUs. However, this is not a change that guarantees improved frame rates across all games. It's a change that reduces selection mistakes that were occurring in cgroup-based environments, with short-lived tasks, and with asymmetric core configurations.
GPU memory allocation priorities are also shifting. The device memory cgroup added in 7.3 provides a kernel mechanism to protect the VRAM used by a foreground game, moving background application data to system memory first when necessary. On discrete GPUs with limited VRAM, this could reduce the stuttering and frame-time instability caused by indiscriminate eviction.
However, the kernel does not automatically determine which application is in the foreground. At this stage, dmemcg-booster enables the feature at the cgroup hierarchy level, while KDE's accessibility features or the new Gamescope designate what gets protected. A formal systemd interface has yet to be built. Upgrading to Linux 7.3 alone won't complete the low-VRAM fix—implementation on the distribution and desktop side is still required.
The word "included" covers four different states of introduction
Changes described as "included" in Linux 7.3 fall into four distinct states of introduction: performance fixes that take effect immediately under specific conditions; kernel mechanisms that require userspace support; initial hardware bring-up that serves as a boot foundation; and removal or deprecation. This classification isn't a mutually exclusive spec breakdown—changes spanning multiple states are placed under whichever condition should be checked first upon adoption. Listing feature names side by side erases this distinction.
| Introduction state | Example in Linux 7.3 | What users should check |
|---|---|---|
| Takes effect immediately on a specific path | Btrfs direct I/O, fsync, and lock-contention fixes | Whether your own workload passes through the affected path, and whether there are regressions |
| Requires userspace cooperation | Device memory cgroup | dmemcg-booster, KDE or Gamescope, and a supported GPU |
| Initial support as a boot foundation | Devicetree for Apple M3 Pro, M3 Max, M3 Ultra | Driver maturity for everyday use, including GPU acceleration |
| Removal or deprecation | EFS, FreeVxFS, legacy 32-bit Arm platforms | Whether old media, hardware, or internal images still depend on them |
The addition of support for Apple's higher-end M3 SoCs centers on the devicetree that lets the kernel recognize the device configuration and proceed with boot. That's still a distance away from completed everyday usability, including GPU acceleration. Meanwhile, EFS, FreeVxFS, and cachefiles' on-demand mode—three subsystems totaling roughly 5,000 lines—have actually been removed. Legacy 32-bit Arm platforms are being deprecated in 7.3, but their removal is still down the road.
This breakdown shows the Linux kernel's development moving in two directions at once: embracing next-generation GPUs and SoCs at an early stage, while stripping out unmaintained old code and rarely used formats. Since the work of adding new features and the work of shrinking the testing surface both draw from the same maintenance budget, these two efforts aren't separate stories—they're connected.
If you're trying Linux 7.3, measure regressions, not averages
Linux 7.3-rc1 is worth trying now for users who can reproduce the same workloads targeted by the improvements and who can report issues. For Btrfs, separate direct I/O testing from actual database workloads; for gaming setups, measure CPU scheduling and VRAM eviction separately. On higher-end Apple M3 machines, distinguish between simply being able to boot and having the devices needed for daily work actually function.
Users deciding whether to move to the stable release should confirm, by the final 7.3 release candidate, whether regressions have disappeared under their own workload, whether their distribution has incorporated the necessary userspace tools, and whether any old formats they depend on are slated for removal. Once those conditions are met, Linux 7.3 will matter not as a release with a long feature list, but as an update that resolves concrete I/O and resource-allocation bottlenecks.