You cannot put code into Linux yourself. It goes in only if the maintainer who owns that part of the kernel takes it, and that happens on a public mailing list where a refusal is as visible as an acceptance and the archive keeps every revision, including the ones that were wrong. Nothing in that thread counts except the patch. What they took is below, each line naming the maintainer who applied it: memory protection for the chip's AI accelerator, support for two new boards, and a fix to Ethernet timing.

iommu/rockchip

Two fixes for the RK3576 NPU's IOMMU instancesApplied by Will Deacon, IOMMU co-maintainer. Mainline, v7.3-rc1.

arm64 dts

ArmSoM CM5 and CM5-IO board supportApplied by Heiko Stübner, Rockchip SoC maintainer; Acked-by Krzysztof Kozlowski. Mainline, v7.3-rc1.

net/phy motorcomm

Reference clock for the YT8521 PHYApplied by Jakub Kicinski, net maintainer; Reviewed-by Andrew Lunn. Mainline, v7.3-rc1.

Firmware is the code that runs before the operating system starts. A trusted execution environment is a sealed-off part of the chip that keeps holding keys and fingerprints even if the operating system itself is broken into. This chip had no such environment until the port below, which the project that maintains the standard reviewed and merged. The rest is secure boot, the same kind of isolation on a second chip, and firmware that boots an ordinary Fedora desktop.

OP-TEE

First RK3576 platform portMerged. Reviewed-by Cherry Embedded Solutions; run by Flipper Devices on Flipper One, 35,681 subtests, zero failures.

Trusted Firmware-A

Redundant secure-boot override removedMerged. Reviewed by engineers at STMicroelectronics, Arm and Rockchip.

RP2350

TrustZone-M isolation exampleSecureFault verified over SWD rather than trusted from source.

UEFI

edk2-rk3576Boots a mainline Fedora 44 desktop; pinned on the Radxa forum by a Radxa engineer.

Nothing on this card has been accepted yet. The first item is a driver that would let a stock Linux use this chip's AI accelerator; it is on its thirteenth public revision, and everyone named can still refuse it. Unfinished work is listed because leaving it out would make the record look tidier than it is.

Linux, accel/rocket

RK3576 NPU enablement, 14 patches, v13Reviewed-by ×5, Acked-by ×2, Tested-by ×5, four of them on RK3588. The driver's author has said he will review it.

Mesa

RK3576 in the Teflon delegate, !43804The bare minimum the maintainer asked for first.

Linux, media

RFC V4L2 driver for the VEPU510 encoderReviewed by Heiko Stübner and Nicolas Dufresne; redirected to the DRM/Vulkan-Video split on their advice.

Nothing here is done. Each line says where the work stands today and what finishing it would mean, so the distance between the two is stated rather than left to the reader.

RK3576 NPU

Into the accel treeThe 14-patch series is at v13; the driver's author has said he will review it. The goal is for it to be applied, so the RK3576 NPU works on a stock kernel.

Mesa

Depthwise, pointwise and image-input convolutions!43804 covers one regular convolution on purpose. The layer types MobileNet is actually built from follow in later series.

charsiu

An open runtime that matches or beats the vendor'sNine models already run end to end on the mainline driver with output identical to the CPU reference. The goal is performance at or above the closed vendor stack, with no vendor code in the path.

Video encoder

Rebase the VEPU510 driver on the DRM/Vulkan-Video splitThe stateful V4L2 interface was abandoned on the maintainers' advice; the driver returns once the RK3588 split it will sit on has landed.

UEFI

A complete UEFI for the RK3576edk2-rk3576 boots a mainline Fedora 44 desktop today. The goal is a full, polished firmware: every board it builds for verified, and the platform support upstreamable.

A DOI is a permanent address for a document, and once it is issued the version behind it cannot be quietly changed. Version 1 of this preprint got its central result wrong, so version 2 withdraws it on the first page and explains the cause. Five findings from the same work were later confirmed on a different chip by people with no contact with this project.

Preprint

RK3576 NPU bring-up, version 2Zenodo DOI. Version 1's central result is withdrawn on the first page of version 2, with the cause.

Cross-checked

Five register-model findings confirmed on RK3588By an independent reverse-engineering effort with no contact with this project.

Coverage

Collabora, Collabora's RK3576 tracker, CNX SoftwareThe Rockchip kernel maintainer commented on Collabora's summary.

Writing a tool proves nothing on its own. The test is whether people who have never spoken to you package it, document it and depend on it. One of these was packaged into Arch Linux by a maintainer I have never contacted, one has its own page in the board vendor's documentation, and one is the subject of that vendor's tutorial.

SoC-Consistency

Static analysis of devicetree power treesPyPI; packaged into Arch Linux by an unaffiliated maintainer.

RKDevelopTool-GUI

Rockchip flashing front endArch AUR; its own page in Radxa's documentation.

kiln

Local LLM inference on the RK3576 NPUSubject of a Radxa tutorial.

charsiu

LLM runtime on the mainline NPU driverNine models run end to end at four bits, output identical to the CPU reference.

Notes written while the work was happening rather than after it succeeded. They include the wrong turns, because a post that records only the version that worked is not much use to the next person.

17 Jul 2026

ArmSoM Noticed the Work, and the CM5 Device Tree Hits v4

13 Jul 2026

Same Chip, Two Walls: TrustZone-M vs RISC-V PMP on the RP2350

15 Jun 2026

Bringing Up the RK3576 NPU on Mainline Linux

Every claim on this site can be checked without taking my word for it. The links lead to servers I do not control: git.kernel.org, the kernel mailing list archive, and Zenodo.

Evidence appendix

Every claim, each with a link to the merged commit, review thread or DOI. Open the appendix

Kernel commits

Linked to git.kernel.org and to the GitHub mirror of Linus's tree.

Names

Kernel patches carry a legal name, so kernel work appears as Jiaxing Hu. Everywhere else I go by gahingwoo, which is also the name to search.

Addresses

Kernel patches and every commit linked from this site carry gahing@gahingwoo.com. huhuvmb88@gmail.com is my school and university account and is attached to my student ID. Both reach me; the first is the one that appears in the public record.