Metal acceleration for the ASRock BC-250 on macOS. The BC-250's GPU (AMD Cyan Skillfish: GC 10.1.3, DCN 2.0.1) is close enough to Navi 10 that Apple's own AMDRadeonX6000 drivers can run it, with a lot of patching. MetalCyan is the Lilu plugin that does that patching. It started as a fork of NootedRed.

For a complete EFI, see BC-250-Hackintosh-OpenCore.

• Metal 3, full acceleration (WindowServer, Safari, Firefox, Metal apps)
• 4K at 60 Hz
• All 4 GB of VRAM usable and CPU-visible
• GPU clock, temperature and load in Activity Monitor, iStat and similar tools
• 40 CU unlock and GPU/CPU overclocking (boot-args, below)

• Hardware video decode/encode. VCN can't be used on this chip, so macOS decodes in software. 4K video plays fine in Firefox.
• HDMI/DP audio. The GPU's audio function isn't set up yet. Safari and the TV app refuse to play video without an audio output device; a virtual one (BlackHole etc.) gets around it.
• GPU recovery. Apple's reset path is Navi 10's and hangs this GPU, so it's blocked. If the GPU hangs, the screen freezes until you reboot. It hasn't happened since the VRAM fixes, but there's no way back from it.
• Shutdown/restart. WindowServer panics on the way down (display_mode_did_change ... returns false). It doesn't affect the next boot. Not fixed yet.
• Sleep. Not tested.

• macOS Tahoe 26.7.1, MacPro7,1 SMBIOS.
• BIOS: UMA frame buffer (VRAM) at 4 GB. With the default 512 MB the GPU runs out of memory and freezes on a green screen.
• Lilu 1.7 or newer, loaded before MetalCyan.
• No other GPU kext (WhateverGreen, NootedRed, NootRX).

Download MetalCyan-1.0.0-RELEASE.zip from Releases (or build it, below). Copy MetalCyan.kext to EFI/OC/Kexts and add it to Kernel > Add after Lilu. The EFI repo uses npci=0x3000 in boot-args as well.

Clocks and voltages are applied 60 seconds after boot, so a bad setting can't stop the machine from booting: you can always get to the desktop and take it out again. If the SMU stops responding (telemetry frozen, CPU stuck at one clock), remove the setting and power off for 10 seconds. A restart isn't enough.

What one board runs daily, with the 8-core unlock: bc250cu=40 bc250gfxmhz=2000 bc250gfxmv=1080 bc250cpumhz=4000 bc250cpuvmax=1300.

• sysctl -n debug.bc250.log: MetalCyan's own log since boot. Only errors and warnings, plus the SMU tuning lines. On a normal boot you'll also see createPspDirectory call not found and PSP: firmware load (handle 10) not done: 2. Both are harmless.
• sysctl -n debug.bc250.smu: live SMU telemetry (GPU clock and voltage, CPU voltage, Tctl, per-core clocks).

Apple's driver is three kexts, and each needs something different.

Framebuffer (AMDRadeonX6000Framebuffer). It gets a BC-250 entry in its ASIC tables and three fixes for DCN 2.0.1:

• Scanout and cursor addresses go out as the carve-out's system address, since DCN 2.0.1 has no DCN_VM aperture.
• The display clock dividers are rescaled for the BC-250's DENTIST VCO.
• MPC pipe splits are stopped: DCN 2.0.1 has 4 pipes, and Navi 10's DAL splits onto pipe 5.

HWServices / HWLibs (TTL, CAIL). The device tables and IP version gates accept the BC-250. The SMU, VCN, JPEG, DMCU and MES blocks are stubbed: Linux either leaves them alone on this chip (VCN and JPEG stay powered off, there's no DMCU, GFX10 runs without MES) or drives them with code Apple doesn't have (SMU 11.0.8). PSP loads the cyan_skillfish2 microcode (bundled from linux-firmware) instead of Navi 10's, and MetalCyan waits for those loads to finish before the RLC starts. GVM, GC and SDMA are programmed the way Linux's amdgpu does it on Cyan Skillfish:

• the FB offset is added to page-table and fault addresses
• Cyan Skillfish golden settings
• Linux's RLC start sequence
• no clock or power gating

Accelerator (AMDRadeonX6000).

• VM: page tables, PTEs and VM programs pointing into VRAM get the carve-out's physical address. TLB invalidations are sent in a form this GPU acknowledges, both over MMIO and in the SDMA/PM4 rings. The fault controls match Linux.
• KIQ: SET_RESOURCES is sent as Linux sends it.
• Compute: compute work runs on the GFX ring. GFX1013's MEC doesn't execute dispatches reliably, and Linux doesn't use it either.
• SDMA1: its trap enable and completion events are fixed up so its work doesn't sit unnoticed.
• VRAM: all 4 GB is made CPU-visible, and the reuse cache that leaked VRAM until the screen went green is off.
• Blocked: Navi 10's reset and hang-dump paths, and VCN's properties, so macOS doesn't try a hardware decoder that isn't there.

SMU. PowerPlay is blocked: Apple's SMU code speaks Navi 10's message set, which the BC-250's SMU 11.0.8 reads differently. MetalCyan talks to the SMU directly through the same mailbox the BC-250 community tools use, for telemetry and the tuning boot-args.

git clone --recursive https://github.com/amethyst8118/MetalCyan.git
cd MetalCyan
make CONFIG=Release zip

Needs clang 19 or newer (for #embed) and ld64. On Linux that's cctools-port; on macOS, point LD64/STRIP at Xcode's. The zip ends up in build-linux/Release/.

• ChefKiss for NootedRed, which this is built on.
• Acidanthera for Lilu and MacKernelSDK.
• Linux's amdgpu driver, the reference for nearly every hardware fix here.
• linux-firmware for the cyan_skillfish2 microcode (MetalCyan/Firmware, under LICENSE.amdgpu).
• The SMU tooling, all of which the SMU code follows (facts only, no code copied from the 40 CU unlock):
  • bc250-collective/bc250_smu_oc: its bc250_smu package for the mailbox, and the CPU voltage model
  • rw-r-r-0644/bc250-core-unlock for the core unlock
  • rw-r-r-0644/bc250-smu-unlock for the SMU firmware patches the EFI's unlock driver applies
  • duggasco/bc250-40cu-unlock for the 40 CU unlock

Thou Shalt Not Profit License 1.5, same as NootedRed. See LICENSE. The firmware in MetalCyan/Firmware is AMD's, under LICENSE.amdgpu.