Restore, verify, and use both AMD FirePro GPUs in the cylindrical 2013 Mac Pro on Linux.
The Mac Pro (Late 2013) always shipped with two AMD GPUs: dual FirePro D300, D500, or D700 cards. Some Linux installations add supergfxctl, a graphics-switching utility intended for laptops with an integrated GPU and a discrete GPU. On the Mac Pro, it can mistake one FirePro for a switchable dGPU and remove it from PCI when configured in Integrated mode.
This project safely disables that behavior, attempts to restore the removed card live, verifies both GPUs, and includes practical dual-GPU workloads. Some systems may require one reboot after applying the fix.
lspcishows only one AMD display controller after login.- Only one
/dev/dri/renderD*device is present. - The boot log initially shows two GPUs, followed by messages such as:
supergfxd: do_mode_setup_tasks(mode:Integrated, ...)
amdgpu 0000:02:00.0: finishing device.
supergfxd: Removed .../0000:02:00.0
git clone https://github.com/microhustlestack/macpro6-1-linux-dual-gpu.git
cd macpro6-1-linux-dual-gpu
./bin/macpro61-gpu status
./bin/macpro61-gpu fix --dry-run
sudo ./bin/macpro61-gpu fix
./bin/macpro61-gpu testThe fix command:
- Refuses to run unless DMI identifies the machine as
MacPro6,1(unless--forceis explicitly supplied). - Stops and disables
supergfxd.serviceif installed. - Masks the service so a future package update cannot silently remove a GPU again.
- Rescans PCI to restore the missing FirePro immediately.
- Verifies two Apple AMD devices, two
amdgpubindings, and two DRM render nodes.
It does not uninstall packages or modify kernel parameters.
The utility can run directly from the repository. To install it system-wide:
sudo make install
macpro61-gpu statusTo remove the installed files:
sudo make uninstallmacpro61-gpu status
macpro61-gpu fix [--dry-run] [--force]
macpro61-gpu test [--frames N]
macpro61-gpu restore-supergfxd
Only fix and restore-supergfxd require root. The test uses FFmpeg to exercise each H.264 hardware encoder independently and then both simultaneously. If installed, vainfo and vulkaninfo are also checked per PCI device.
Recommended packages on Arch Linux:
sudo pacman -S ffmpeg mesa vulkan-radeon libva-utils vulkan-tools radeontopEquivalent package names vary by distribution.
The initial release was validated on a MacPro6,1 with dual FirePro D500 GPUs, Linux 7.2, Mesa 26.2, and the amdgpu driver:
| Test | GPU at 02:00.0 |
GPU at 06:00.0 |
|---|---|---|
| Independent 1080p60 H.264 encode | ~80 fps | ~80 fps |
| Concurrent 1080p60 H.264 encode | ~84 fps | ~85 fps |
| Vulkan rendering | Passed | Passed |
| VRAM | 3 GiB | 3 GiB |
| Temperature after test | 53 C | 58 C |
Concurrent aggregate encode throughput was approximately 169 fps. No GPU faults, resets, hangs, or PCIe errors were logged.
Validated on a second MacPro6,1 with dual FirePro D300 GPUs (1002:679E, Apple subsystem 106B:0126), Linux 7.2.5-omarchy, Mesa 26.2.2, and the amdgpu driver, running the same 300-frame 1080p60 test:
| Test | GPU at 02:00.0 |
GPU at 06:00.0 |
|---|---|---|
| Independent 1080p60 H.264 encode | ~12 fps | ~79 fps |
| Concurrent 1080p60 H.264 encode | ~12 fps | ~85 fps |
| Vulkan rendering | Not checked | Not checked |
| VRAM | 3 GiB | 3 GiB |
| Temperature after test | 55 C | 59 C |
Both cards passed every encode test with no faults, resets, hangs, or PCIe errors. On this unit the 02:00.0 VCE engine sustained a fixed low rate of roughly 12 fps regardless of resolution or test pattern, while 06:00.0 matched the D500 throughput; the cause appears to be an unresolved VCE clock cap on that card, and D300 owners should expect per-card variance. vainfo and vulkaninfo were not installed on this system, so VA-API capability and Vulkan were not exercised.
Performance varies with CPU speed, source decoding, cooling, Mesa, kernel, and GPU model.
Each FirePro has its own H.264 VCE engine. Process two independent videos at once, one per render node:
./examples/dual-transcode.sh ./encoded video-one.mkv video-two.mkv video-three.mkvThe script runs at most two jobs concurrently and assigns them round-robin to the two PCI-addressed render nodes.
Keep the display-connected GPU responsive while running video work on the headless card:
ffmpeg -vaapi_device /dev/dri/by-path/pci-0000:02:00.0-render \
-i input.mkv -vf 'format=nv12,hwupload' \
-c:v h264_vaapi -qp 22 -c:a aac -b:a 192k output.mp4PCI addresses can differ. Always use macpro61-gpu status and /dev/dri/by-path/ rather than assuming card numbers.
Mesa supports selecting a GPU by PCI address:
DRI_PRIME=pci-0000_02_00_0! vulkaninfo --summary
DRI_PRIME=pci-0000_06_00_0! vulkaninfo --summaryApplications must support explicit multi-GPU operation or be launched as separate processes. The GPUs' VRAM is not pooled: dual 3 GiB D500s do not become one 6 GiB device.
More examples and limitations are documented in Use Cases.
sudo macpro61-gpu restore-supergfxdThis only removes the service mask. It intentionally does not enable or start supergfxd, because starting it in Integrated mode can immediately remove one GPU again.
The hardware detection covers all Apple GPU options for MacPro6,1 without relying on individual AMD device IDs:
- Dual FirePro D300 with 2 GiB each
- Dual FirePro D500 with 3 GiB each
- Dual FirePro D700 with 6 GiB each
The D300 and D500 configurations have been tested directly. Reports and patches for D700 systems are welcome.
