Video summary: this article is a detailed, in-depth summary of the video “Linux 7.3 Review: AI Chaos, MASSIVE Performance, GPU, CPU, and Rust Upgrades”, from the channel SavvyNik, combined with Linus Torvalds' official announcement closing the merge window.
Video URL: https://www.youtube.com/watch?v=tQ5adDeOGb0
The merge window for Linux kernel 7.3 has closed, and 7.3-rc1 is out. It is a packed cycle: gains of up to 5x in Btrfs, latency reductions of 90% in memory management, Nvidia's Rust driver maturing, initial support for Apple M3 chips — and, behind the scenes, maintainers drowning in AI-generated patches. Let's take it piece by piece.
Linus' official announcement
In his merge-window closing email, Torvalds summed up the mood (freely translated):
“Two weeks have passed, it is Sunday afternoon, and the merge window has closed. Nothing really stands out — except that it is big. […] It is not the biggest rc1 we have ever had, but it is certainly up there in commit count, and I think that no longer surprises anyone. […] This is another of those releases with a large dump of AMD GPU registers — this time the DCN6 headers. […] That AMD code and those headers account for about a third of the entire rc1 patch. If you mask out that block, things look pretty normal.”
And with his usual candor, Linus admitted making his own life harder: “I did something stupid by upgrading the system in the middle of the window and had to track down a problem it caused — but it was entirely my fault”. Even the creator cannot avoid breaking his own Linux during an upgrade. Technical context: DCN6 is the new generation of AMD's Display Core Next (the GPU component that handles monitors and timings), and the headers contain thousands of C register definitions — hence the volume.
Nearly 2.000 CVEs and AI chaos
Greg Kroah-Hartman brought the alarming number: after 7.2 exceeded 1.500 CVEs, 7.3 is approaching 2.000 CVEs per release. Another problem is growing alongside it: maintainers are being buried in low-priority submissions — cleanups, tiny fixes, and “clarifications” generated by AI/LLMs. The networking team was blunt: “we are completely overwhelmed” — with the irony that the lifeline was “LLM budget secured, thanks Meta”: they are fighting AI with AI to triage more than 630 networking patches in the cycle. All of this in a kernel approaching 40 million lines of code — of which 6 to 7 million are the amdgpu driver alone.
Btrfs: up to 5x faster
The cycle's big performance win: Btrfs now uses iomap as a bounce buffer for direct I/O instead of falling back to slow buffered I/O. Workloads that ran at ~50% of the theoretical maximum moved to ~95% — effectively double the throughput, and some workloads improve by 3 to 5x. Important fixes also arrived: folio protection during writeback in zoned mode (preventing the system from mounting read-only under transient overcommit), and a deadlock between defragmentation and delayed-allocation reservations.
Scheduler: smoother games
The highlight is “flatten the pick” in cgroup scheduling, bringing CFS closer to a single-queue model — tests on older CPUs showed better gaming performance. Also included: lower latency for short-slice tasks, scheduling fixes for clusters with asymmetric capacity, and a preference for fully idle cores in NOHZ balancing.
Gaming: native Steam Controller support in the kernel
The new Steam Controller now works through the kernel's standard HID stack — without depending on Steam or SDL: it is a regular Linux input device. HID also gained support for the MSI Claw gamepad, the microphone mute status of the HyperX Quad 2 and the programmable buttons of Logitech HID++ 2.0.
Apple Silicon: M3 support arrives
Included are the device trees for M3 Pro, M3 Max, and M3 Ultra — the foundation for the kernel to recognize and begin booting these machines (this is initial support, not full functionality). As a bonus: hardware-monitoring sensors for M1/M2 and adjustments that save ~1 watt on M1/M2 Pro/Max/Ultra. Yes, we are playing catch-up — Apple already sells the Mac Mini M6 — but that is the cost of reverse engineering.
Memory: three latencies crushed
- ZRAM/zswap: less lock contention when freeing compressed pages — an operation 1,83x faster on a Raspberry Pi and 1,4x on a 20-core Intel
- KSM (same-page merging): a reverse-mapping operation that held memory for 500–700 ms dropped to less than 2 ms in the worst case
- TLB flush on x86: the CPU waited for other cores to invalidate addresses with preemption disabled — delays of up to 16 ms on a 16-core system are now ~1,5 ms (up to 90% better). Gold for telecom, audio, and trading
- Along the same lines, SMP behavior was fixed to re-enable preemption while waiting for remote CPUs: urgent-task latency fell from ~17 ms to ~1,5 ms
GPUs: AMD, NOVA in Rust, and Nouveau
At AMD: the new memcg-aware VRAM eviction handles video-memory shortages intelligently — work by an engineer at Valve, tested with Indiana Jones: The Great Circle configured for 9 GB on an 8 GB GPU, running smoothly thanks to selective eviction to RAM. Older cards (GFX6/GFX8 era) gain DRM modifiers that unlock modern Wayland and Vulkan, and new hardware gains the display engine DCN 6.0 with 8K fixes.
At Nvidia: the NOVA — the next-generation driver written in Rust, Nouveau's successor — advances with booting of the GSP (the GPU system processor that takes over low-level management), vGPU, TLV firmware, and more Hopper and Blackwell support. And Nouveau gains video decoding through NVDEC for Mesa's NVK Vulkan driver — hardware-accelerated H.264 first, HEVC on the way.
USB-C and Thunderbolt: the end of mystery cables
The kernel can now read the capabilities of the cable itself for USB-C: whether it supports DisplayPort/Thunderbolt alt-mode, whether it is active or passive, and which speeds it can handle. Two cables that look identical can be a charge-and-USB-2.0-only cable and a full Thunderbolt cable — Linux now knows the difference, which should reduce mysterious dock and monitor failures. Also arriving is busy polling for USB4: lower latency for direct transfers between machines, at the cost of more CPU.
Filesystems: iomap, FUSE, and even Time Machine
- iomap completed its migration to the iterator model: fewer indirect calls and, on PCIe Gen 5 NVMe, an increase from 1,92 to 2,19 million IOPS (+4–5%)
- FUSE gains io_uring buffer pools and zero-copy: less copying between kernel and userspace, narrowing the gap with in-kernel filesystems
- KSMBD (the kernel SMB server) now speaks Apple's proprietary extensions: a Linux share can be a backup destination for Time Machine
- The foundation (DAX/NVDIMM) arrives for FAMFS, Micron's filesystem for fabric-attached memory — shared pools of hundreds of terabytes, targeting AI
- Out go the fossils EFS and freevxfs, and FAT fixes a classic bug: names over 255 bytes used to be silently truncated; now the operation fails correctly
Firmware, power, and security
- EFI with a timeout: a stuck runtime-service call no longer freezes the system — after 120 seconds the kernel marks the path as broken and reports it
- Power: more accurate Intel P-state during idle; on AMD, dynamic per-core EPP decides power versus performance core by core — great for laptop battery life
- AF_ALG at the end of the line: the new
af_alg_restrictexposes only the still-needed algorithms by default, reducing the attack surface of the socket-based cryptography API - KCFI (control-flow integrity) rewritten to avoid penalizing performance on older Intel CPUs
- Security cleanup: the SGI GRU drivers (2012 hardware, ~7.000 unused lines) were deleted, and a large batch of legacy 32-bit ARM (pre-device-tree board files, old EABIs) entered deprecation — a smaller kernel is a more defensible kernel
Rust spreads, and Linux goes to space
Rust in the kernel takes two portability steps: preparation for the experimental backend rustc_codegen_gcc (unlocking architectures only GCC supports, such as m68k and Alpha, and reducing LLVM dependency) and support for PowerPC — 32-bit big-endian and 64-bit little-endian join x86-64, ARM64, RISC-V, and LoongArch. There is even a space-related note: initial support arrives for Intel Starfire, a radiation-tolerant SoC derived from Panther Lake, aimed at spacecraft and aerospace — with memory-error correction included. Networking rounds out the list: PHYs Realtek RTL8261C/D at 10 GbE now work with the stock kernel, without a vendor driver.
The bottom line
7.3 is the kind of release without a single headline, but with dozens of accumulated wins: flying Btrfs performance, crushed memory latencies, gaming increasingly a first-class citizen (thanks, Valve), Nvidia's Rust driver becoming reality, and legacy code cleanup on track. rc1 is finalized; now come ~7 weeks of release candidates before the final version. If you want to follow closely, SavvyNik's video details every patch with source links — it is worth subscribing.