A native macOS display driver for AMD Navi 32 (RDNA 3) GPUs.
Starting with the Radeon RX 7800 XT (1002:747E): native display, Vulkan through RADV, and Metal on the way.
Research first, hardware claims backed by evidence.
For developers only. This is a research prototype, not a driver you can install and use. It needs OpenCore, SIP turned off and boot-args, it writes to the GPU from the kernel, and it can hang or crash your machine. Run it only on a system whose display does not depend on this card, with a backup EFI. Vulkan (Mesa's RADV, through a bridge to the kext) runs on the card in test programs. Metal is at an early stage: on the development machine WindowServer composes the desktop on the GPU through our plug-in, and since 2026-11-21 without the visual glitches it had.
Join the community: Discord for questions, test results and hardware support (RX 7600, RX 7700 XT, APUs), and the issues for the work itself. Testers with a Linux or Hackintosh machine are welcome: the first step, a read-only capture, is in docs/ADDING_A_GPU.md.
Goal: a real framebuffer on the card, driven by our own kext (DCN 3.2 display engine), with no GOP, VESA or firmware framebuffer in the way, and then acceleration: RADV (Vulkan) on top of the kext, which runs compute and 3D tests today, and Metal on top of RADV, which has started (the desktop is composed on the GPU on the development machine) and is where we need help.
Pre-alpha, 2026-11-10. Display (2026-10-28): the macOS Sequoia desktop is shown on a real RX 7800 XT, over DisplayPort, through our own kext. The kext loads the card's PSP firmware chain and our own DMCUB display firmware, trains the DisplayPort link, lights the pipe in a mode read from the monitor's EDID (tested from 1024x768 to 1920x1080 at 144 Hz) and scans out a linear framebuffer in VRAM that WindowServer draws into; the resolution and refresh rate can be changed from System Settings while it runs. The picture does not come from the firmware's GOP/VESA framebuffer: without the kext the card's output stays dark.
Acceleration (2026-11-02): the kext loads the GFX, MES, SDMA, IMU and RLC firmware through the PSP, starts the MES and maps a compute queue and a graphics queue through it, gives each user-space connection its own GPU address space, and manages 16 GB of VRAM. Mesa's RADV runs on top of it on macOS (a patch set and a bridge in mesa-bridge/, no libdrm): vulkaninfo lists the RX 7800 XT, and test programs run compute dispatches (a burst of 20000 submits, ring saturation, 20000 dependent barriers, five processes at once), draw triangles, and render a scene with a depth buffer, a sampled texture and 4x multisampling that is shown on the monitor through the display surface. A hung queue is reset by the kernel (compute and graphics). All of this is test programs run as administrator on a development machine; the numbers and what is not done are in docs/COMPAT.md and docs/M5_MESA_BUILD.md.
Desktop on the GPU (2026-11-07 to 2026-11-10): with the accelerator kext (RedNaviAccel2) and the Metal plug-in (RedNaviMTL, Metal on RADV) installed, WindowServer picks its Metal compositor and the desktop is composed on the RX 7800 XT, at up to 144 Hz, with a hardware cursor (DCN cursor plane). This runs on one machine and is not stable: animations can leave old pixels on the screen, frames tear (each frame is copied into the surface the display is scanning), about one frame in ten is late, and a shader is compiled the first time an effect appears. The state, the measurements and the open problems are in docs/M6_DISPLAY_ACCEL.md.
The glitches are gone (2026-11-21). For two weeks the GPU-composed desktop had grey rectangles in window title bars and browser tab strips, and no change to texture uploads, bindings or surface syncing removed them. The cause was the order of blended draws inside one render pass: on RDNA 3 with our RADV bridge, a blended draw could read the colour target before the draws in front of it had written it, so rectangles of the screen came out half drawn. darrenintr found it on an RX 7600 with an offscreen CoreAnimation test compared pixel by pixel against Core Graphics (23 bad frames in 1200 before, 0 after). The same one-barrier fix, ported to the RX 7800 XT, left the desktop clean: Brave, the Claude app, windows and animations all render without a single grey block. Why the hardware needs a barrier that Vulkan's rasterization order should make unnecessary is not known yet; it is off with RN_VK_BLENDSYNC=0 for anyone who wants to dig.
What this is not: Metal is not complete (no depth or stencil, no multisampling, few attachment formats) and nothing but the system compositor has been run on it in daily use; for Vulkan the only window system integration is a swapchain on a headless surface that owns the whole display (one application at a time, no windows), and there is no video, no power management of the GPU, no interrupts to the CPU (every wait polls). Modes come from the EDID, but only on one DisplayPort connector, and the display-mode bandwidth values are borrowed from a 1080p60 capture, so unusual timings may underflow. There is no clock or power management (SMU), hot-plug, audio or HDMI, and only one DisplayPort connector has been tested. A passive DisplayPort to HDMI cable gives no picture: it needs the connector switched to TMDS (DP++), which is not written. It needs OpenCore, and the framebuffer kext needs SIP off (details below). It can hang the machine. There is no release.
| Milestone | What | State |
|---|---|---|
| M0 | Repo, CI (macOS runner), build scripts | Done, CI green (Debug and Release) |
| M1 | Probe kext: matches 1002:747E, maps the register BAR, reads the IP discovery table |
Done: ran on a real RX 7800 XT, versions recorded in docs/ARCHITECTURE.md |
| M1.5 | Probe stage 2: read-only PSP and SMU mailbox registers | Done: PSP bootloader ready, sOS not running, SMU alive |
| M2a | PSP bring-up and DMCUB load, as opt-in probe stages (PSP firmware chain, secure OS, command ring, LOAD_TOC, SETUP_TMR, DMCUB via the PSP) |
Done on a real RX 7800 XT: the secure OS answers commands, and our DMCUB firmware runs and answers mailbox commands (see docs/PSP_BRINGUP.md, docs/DMUB_BRINGUP.md) |
| M2b | SMU 13 bring-up, and the LinuxKPI-style compatibility layer that replaces the probe's hand-written register code | Started: the compat layer exists with host tests (docs/COMPAT.md); the SMU answers messages after its firmware is loaded through the PSP. Clock and power management on top of that is not written yet |
| M3 | Native framebuffer: DCN 3.2 display engine + IOFramebuffer |
Works on one DisplayPort connector: link training, a pipe state replayed from a Linux capture with the mode-dependent registers recomputed (OTG, DP stream, DTO, plane sizes, DISPCLK), the monitor's EDID modes, and live mode changes through RedNaviFB (see docs/DC_PLAN.md, docs/DP_BRINGUP.md). Missing: more connectors, a real DC port instead of a replay, exact bandwidth (DML) values |
| M4 | GFX11, MES, SDMA 6, GPU VM, user client | Done for compute and graphics (2026-11-01 and 02): PSP-loaded firmware, RLC boot, GART and VRAM page tables, the MES maps a compute queue and the graphics queue (ADD_QUEUE), per-connection GPU address spaces (VMIDs 1 to 7, 4-level tables), VRAM allocator (13.5 GB device-local, 78 MB CPU-visible), sequence-numbered fences per queue, DRM-style syncobjs with timelines, queue reset and a 8 s watchdog on both queues, a user client with a versioned struct interface (tools/rn_test.c tests it). Details in docs/COMPAT.md. Missing: SDMA through the MES, MSI interrupts to the CPU, GC power management (GFXOFF is held off), the clear-state buffer and golden registers of the RLC, buffers above 1 GB, clearing of device-local buffers before reuse (so the client must be an administrator) |
| M5 | RADV (Mesa) running on macOS through an IOKit backend for Mesa's ac_drm_* layer (plan: docs/M5_PLAN.md, build and results: docs/M5_MESA_BUILD.md) |
Works in test programs (2026-11-02): RADV builds on macOS (Mesa bd830ce, -Dvulkan-drivers=amd, ACO), enumerates the card, compiles shaders and runs compute, 3D and multisampled rendering, checked pixel by pixel and on the monitor. Missing: window system integration, many Vulkan features (stencil, blending, filtering, tessellation, ray tracing, video are untested), sparse resources, external memory, performance work |
| M6 | Metal plugin on top of RADV + accelerator kext for WindowServer (plan: docs/M6_PLAN.md, kernel side: docs/M6_ACCELERATOR_STUDY.md, plug-in side: docs/M6_METAL_ON_RADV.md) | In progress (2026-11-07): RedNaviAccel2 lets Apple's own IOAcceleratorFamily2 start on the card and macOS Metal lists the RX 7800 XT as a device (opt-in by boot-arg rednavi_accel2=3). The Metal plug-in (RedNaviMTL/) runs, through the official Metal API on RADV and verified on the card: compute kernels compiled by Apple's metal (AIR translated to SPIR-V by a separate LGPL component), buffers with offsets, textures and samplers, render pipelines with vertex descriptors ([[stage_in]]), draws (indexed or not, culling, alpha blending) into textures, passes that only clear, a blit encoder, and textures made from IOSurfaces (by copy, or over the surface's own memory when it is one linear plane); a render target can be put on the monitor through the page flip. 2026-11-07: WindowServer composes on the GPU for the first time (Metal compositor, CoreAnimation and CoreDisplay's GPU pipe through the plug-in, frames copied into the scanout; transparency seen on the monitor), see docs/M6_DISPLAY_ACCEL.md. Not stable: WindowServer crashes seconds after the mouse moves, frames glitch, WindowServer must be restarted by hand after the accelerator appears. 2026-11-09 and 10: one Vulkan submit per command buffer, zero-copy IOSurfaces, CoreDisplay's scanout surfaces rendered in VRAM, a hardware cursor, the display link at the mode's refresh rate. Old pixels after animations and tearing are open. Not done: depth and stencil, multisampling, several attachment formats, video decode, a distribution of the translator without LLVM tools |
M4 and M5 give you compute and Vulkan on the card but do not change how the desktop feels; a smooth, GPU-composited desktop needs M6, which is under way: the desktop is composed on the GPU on the development machine, and the rendering glitches are fixed; stability and frame pacing are the open work.
macOS has no RDNA 3 support at all and Apple's Metal shader compiler has no GFX11 back end, so there is nothing to impersonate. The approach that has been shown to work for an unsupported GPU is the one used by NullMoth for NVIDIA: host the vendor's open kernel driver inside a kext, run the vendor's Vulkan driver in userspace, translate Metal to Vulkan on top.
For AMD that means: amdgpu in a kext, RADV (ACO) in userspace, Metal -> SPIR-V translation on top. See docs/ARCHITECTURE.md.
ASUS ROG STRIX Z490-G GAMING, i7-10700K, Intel UHD 630 as the primary display, Sapphire RX 7800 XT as the development target (PCIe 4.0 x16). See docs/TEST_MACHINES.md.
The RX 7800 XT is the only board the kext writes to. The RX 7700 XT (same die and device ID, told apart by VRAM size) and the RX 7600 (Navi 33) are in the ASIC table from Linux captures, and the probe stays read-only on them. The RX 7600 has run once on macOS (2026-11-16, on a machine that reports a Mac Pro SMBIOS, thanks to GGBBZZS): the kext loaded, the probe named the card Navi 33 and the IP discovery table it read agrees with the Linux capture in all 39 blocks (docs/logs/navi33/rx7600-macos/). That is a read-only result: no display, no firmware and no register writes on that card yet. The RX 7700 XT has not run on macOS (unverified). Reports from other 7000-series owners are welcome.
Needs macOS with the Xcode command-line tools, or just push to GitHub: .github/workflows/build.yml builds Debug and Release on a macOS runner and uploads the kext.
sh scripts/bootstrap.sh
make DEBUG=1
Load RedNavi.kext on a machine where the 7800 XT is not the display you depend on. By default the probe only reads:
ioreg -rn RedNaviProbe -l -x > probe.txt
python3 tools/decode_discovery.py probe.txt
It publishes the IP discovery table (ip-discovery), PSP/SMU registers (mailbox-regs), MMHUB and DMCUB registers and the VBIOS read from the PCI ROM as properties. IOLog output may not show up in log show (cause unknown), so read the properties.
The stages that write to the GPU (PSP load, ring, TMR, DMCUB load, DisplayPort training, the display replay) are off unless enabled with rednavi_* boot-args, and need firmware headers generated locally from linux-firmware (tools/gen_psp_fw_header.py, tools/gen_dmub_fw_header.py); firmware is not redistributed here. The boot-args are listed in docs/PSP_BRINGUP.md, docs/DMUB_BRINGUP.md and docs/DP_BRINGUP.md. They can hang the machine: use a system whose display does not depend on this card, with a USB copy of your EFI.
Needs boot-args for the acceleration stages (rednavi_mes_step=5 rednavi_gc_init=1 rednavi_gc_awake=1 rednavi_mmhub=1, plus rednavi_gfx_queue=1 for the graphics queue; the full list is in docs/COMPAT.md), the firmware headers for the GC, MES and SDMA files (tools/gen_gc_fw_header.py), and an administrator account. The Mesa side is built from a Mesa checkout with mesa-bridge/patches/0001-radv-darwin-iokit.patch (recipe in docs/M5_MESA_BUILD.md); mesa-bridge/tools/ has the ICD manifest and run scripts. Then vulkaninfo, and the programs in mesa-bridge/test/: vk_dispatch (compute), vk_stress (bursts, barriers, VRAM), vk_draw (a triangle), vk_present (a triangle, scene for depth, texture and 4x MSAA, spin for an animation) and vk_swapchain (the cubes through VK_KHR_swapchain). tools/desktop_test.sh deploys the kext to a test machine over SSH, reboots it and runs the kernel-level tests of tools/rn_test.c.
With rednavi_fb=1 (and rednavi_dcn_replay=14) the probe publishes where the scanout surface is, and a second kext, RedNaviFB.kext, hands it to WindowServer. It is a separate kext because it links against IOGraphicsFamily, which lives in the system kernel collection, so OpenCore cannot inject it. Install it through the auxiliary kernel collection (SIP off, codesign -f -s -, kmutil load -p), as described in docs/DP_BRINGUP.md. If macOS also shows a second, nameless display, that is the stock firmware boot framebuffer, not a second output. Boot-arg -wegnoigpu (WhateverGreen) switches the Intel iGPU off so the card is the only display.
Linux owners (no macOS needed): sudo python3 tools/linux_collect.py from a text console with the monitor lit writes one rednavi_report_*.tar.gz (IP blocks, connectors, amdgpu log lines, the display registers of the connector that is lit; no EDID, no serial numbers) to attach to an issue.
Please post decoded output in an issue, together with your board and BIOS settings (Above 4G Decoding, Resizable BAR).
- Helemen7 — Project founder, kernel display architecture, Metal/RADV translation layer.
- Frame121 — Navi 33 (RX 7600) Linux IP-discovery capture, directory layout parsing, and test evidence.
- Naessah — Strix Halo (Ryzen AI Max+ 395 / Radeon 8060S) RDNA 3.5 live IP-discovery capture, UMA APU architecture, and hardware evidence.
- GGBBZZS — RX 7600 (Gigabyte, Navi 33) Linux captures: IP discovery, firmware answers and the HDMI display and plane registers from a text console, and the project logo; the first macOS run on an RX 7600 (Mac Pro SMBIOS), which confirmed the Navi 33 IP discovery table on macOS.
- darrenintr — The AIR to SPIR-V translator library the Metal plug-in loads, the Navi 33 (RX 7600) desktop work, and the find that ended the glitches of the GPU-composed desktop: blended draws in one pass need an explicit colour barrier on this stack, found with an offscreen CoreAnimation test checked pixel by pixel against Core Graphics. Also the cause of the trails of the minimize animation: RADV set the attribute ring's BIG_PAGE although our page tables use 4 KB pages, which corrupts draws that span more than one NGG workgroup (gone on the RX 7800 XT as well).
- porcelainhorse — RX 7700 XT (Gigabyte, Navi 32) Linux captures: IP discovery, raw DRM info, DisplayPort registers, and linear scanout plane registers from a text console.
We want contributors. Good first areas, each of which will get an issue:
- LinuxKPI-style compatibility layer for XNU (memory, locks, workqueues, dma-fence, firmware loading)
- PSP 13 / SMU 13 bring-up
- DCN 3.2 modeset from the EDID and more connectors (HDMI, second DisplayPort)
- SMU 13 clock and power management
- Hot-plug, HDMI and DP++ adapters, cursor sizes and modes other than the tested one
- Other cards: the path is in docs/ADDING_A_GPU.md; the Navi 32 constants that still have to move behind the ASIC table are listed there (issue #7)
- SDMA through the MES, interrupts (MSI/IH) to the CPU, GC power management (GFXOFF) (finishing M4)
- Vulkan feature coverage on RADV: stencil, blending, filtering and mipmaps, indexed draws, more formats, uniform buffers; bug reports with a test program are as useful as fixes (M5)
- The Metal side (M6): the AIR to SPIR-V translator, depth and stencil, old pixels and tearing in the composed desktop, frame pacing, and hardening of the accelerator kext (the swap chain for Vulkan already works)
- Decoding the IP discovery dump (
tools/)
Please do not copy code from projects with restrictive licences (NullMoth's kexts and plugin are PolyForm Noncommercial). The Metal translator in that project is LGPL-3.0 and may be reused as a separate component.
BSD-3-Clause for the code in this repository. Register definitions and behaviour taken from the Linux amdgpu driver are MIT, see NOTICE.
GPU firmware is not included. It comes from linux-firmware under LICENSE.amdgpu, which allows redistribution in binary form only, with AMD's notice reproduced, and forbids reverse engineering, decompilation and disassembly. That is why firmware is never committed here and each user generates local headers from files they fetch themselves.