moderately experienced linux nerd here but having a tough go with a somewhat unique setup. I am running Arch on a 2013 Mac Pro (6,1). Since the built in HDMI port only supports HDMI 1.4, I would like to use an eGPU (RTX 2070 super) with a Sonnet Egfx Breakaway Box 650 (using this as a HTPC and would like HDMI 2.0). The Sonnet is a Thunderbolt-3 device, so I have a Genuine Apple Thunderbolt 3 to Thunderbolt 2 Adapter. I understand the bandwidth bottleneck with PCIe 3.0 x16 vs Thunderbolt 3/2. However, it seems that others have been able to get this to work – https://schoolofmotion.com/blog/2013-mac-pro-egpu.

I have tried all different combinations of nvidia drivers (open, dkms, 470, recent, etc.) as well as several combinations of kernel paramenters. I believe the root cause is the error “cannot obtain PCI resources” (see below). I tried with an AMD Vega 56 GPU and also had the same issues.

sudo journalctl -b | grep thunderbolt
Aug 08 10:17:26 archbox kernel: ACPI: bus type thunderbolt registered
Aug 08 10:17:26 archbox kernel: thunderbolt 0000:16:00.0: device link creation from 0000:15:00.0 failed
Aug 08 10:17:26 archbox kernel: thunderbolt 0000:5d:00.0: device link creation from 0000:5c:00.0 failed
Aug 08 10:17:26 archbox kernel: thunderbolt 1-1: new device found, vendor=0x8 device=0x4a
Aug 08 10:17:26 archbox kernel: thunderbolt 1-1: Sonnet Technologies, Inc. eGFX Breakaway Box 650 OC
Aug 08 10:17:26 archbox kernel: thunderbolt 0000:a4:00.0: error -EINVAL: cannot obtain PCI resources, aborting
Aug 08 10:17:26 archbox kernel: thunderbolt 0000:a4:00.0: probe with driver thunderbolt failed with error -22

lspci -nnk | grep -iA3 vga | grep -vi amd | grep -v Apple | grep -v "-"
60:00.0 VGA compatible controller [0300]: NVIDIA Corporation TU104 [GeForce RTX 2070 SUPER] [10de:1ec7] (rev a1)
    Subsystem: NVIDIA Corporation Device [10de:13aa]
    Kernel driver in use: nvidia
    Kernel modules: nouveau, nvidia_drm, nvidia

sudo dmesg | grep NVRM
[ 13.543090] NVRM: loading NVIDIA UNIX x86_64 Kernel Module 470.256.02 Thu May 2 14:37:44 UTC 2024
[ 213.720684] NVRM: GPU 0000:60:00.0: RmInitAdapter failed! (0x22:0xffff:667)

sudo journalctl -b | grep nvidia
Aug 08 10:17:26 archbox kernel: nvidia: loading out-of-tree module taints kernel.
Aug 08 10:17:26 archbox kernel: nvidia: module license 'NVIDIA' taints kernel.
Aug 08 10:17:26 archbox kernel: nvidia: module verification failed: signature and/or required key missing - tainting kernel
Aug 08 10:17:26 archbox kernel: nvidia: module license taints kernel.
Aug 08 10:17:26 archbox kernel: nvidia-nvlink: Nvlink Core is being initialized, major device number 237
Aug 08 10:17:26 archbox kernel: nvidia 0000:60:00.0: enabling device (0000 -> 0002)
Aug 08 10:17:26 archbox kernel: nvidia 0000:60:00.0: vgaarb: VGA decodes changed: olddecodes=io+mem,decodes=none:owns=none
Aug 08 10:17:26 archbox kernel: nvidia-modeset: Loading NVIDIA Kernel Mode Setting Driver for UNIX platforms 470.256.02 Thu May 2 14:50:40 UTC 2024
Aug 08 10:17:26 archbox kernel: nvidia_uvm: module uses symbols nvUvmInterfaceDisableAccessCntr from proprietary module nvidia, inheriting taint.
Aug 08 10:17:26 archbox kernel: nvidia-uvm: Loaded the UVM driver, major device number 235.
Aug 08 10:17:26 archbox kernel: [drm] [nvidia-drm] [GPU ID 0x00006000] Loading driver
Aug 08 10:17:26 archbox kernel: [drm] Initialized nvidia-drm 0.0.0 20160202 for 0000:60:00.0 on minor 0
Aug 08 10:17:26 archbox kernel: Modules linked in: wl(POE+) cfg80211 rfkill crypto_user loop nfnetlink ip_tables x_tables hid_logitech_hidpp dm_crypt hid_logitech_dj cbc encrypted_keys trusted asn1_encoder tee dm_mod crct10dif_pclmul crc32_pclmul polyval_clmulni polyval_generic gf128mul ghash_clmulni_intel sha512_ssse3 sha256_ssse3 sha1_ssse3 aesni_intel nvme crypto_simd cryptd nvme_core nvme_auth xhci_pci xhci_pci_renesas nvidia_drm(POE) nvidia_uvm(POE) nvidia_modeset(POE) nvidia(POE) amdgpu amdxcp drm_exec gpu_sched drm_buddy drm_ttm_helper ttm video wmi i2c_algo_bit drm_suballoc_helper drm_display_helper cec hid_generic usbhid tg3 ptp pps_core libphy ext4 crc32c_generic crc32c_intel crc16 mbcache jbd2
Aug 08 10:17:26 archbox kernel: Modules linked in: apple_gmux(+) fjes(-) mac_hid vboxnetflt(OE) vboxnetadp(OE) vboxdrv(OE) coretemp i2c_dev wl(POE) cfg80211 rfkill crypto_user loop nfnetlink ip_tables x_tables hid_logitech_hidpp dm_crypt hid_logitech_dj cbc encrypted_keys trusted asn1_encoder tee dm_mod crct10dif_pclmul crc32_pclmul polyval_clmulni polyval_generic gf128mul ghash_clmulni_intel sha512_ssse3 sha256_ssse3 sha1_ssse3 aesni_intel nvme crypto_simd cryptd nvme_core nvme_auth xhci_pci xhci_pci_renesas nvidia_drm(POE) nvidia_uvm(POE) nvidia_modeset(POE) nvidia(POE) amdgpu amdxcp drm_exec gpu_sched drm_buddy drm_ttm_helper ttm video wmi i2c_algo_bit drm_suballoc_helper drm_display_helper cec hid_generic usbhid tg3 ptp pps_core libphy ext4 crc32c_generic crc32c_intel crc16 mbcache jbd2
Aug 08 10:17:26 archbox kernel: nvidia-gpu 0000:60:00.3: enabling device (0000 -> 0002)
Aug 08 10:17:38 archbox nvidia-persistenced[1848]: Started (1848)
Aug 08 10:17:38 archbox nvidia-persistenced[1848]: device 0000:60:00.0 - failed to open.

What We Tried:
I tried various combinations of NVIDIA drivers, including open-source, DKMS, 470xx, and the most recent drivers. Additionally, I experimented with multiple kernel parameters to troubleshoot the “cannot obtain PCI resources” error. I also attempted using an AMD Vega 56 GPU to see if it would yield different results.

What Was Expected:
I expected the eGPU (RTX 2070 Super) to be detected and function correctly, allowing for HDMI 2.0 output via the Sonnet eGFX Breakaway Box 650. Specifically, I was looking to bypass the limitations of the built-in HDMI 1.4 port on the Mac Pro by using the Thunderbolt-to-eGPU setup.

What Actually Happened:
Despite the various drivers and kernel tweaks, the system consistently failed to obtain PCI resources for the eGPU, leading to errors and the GPU not being recognized or utilized as expected. This issue persisted even when using the AMD Vega 56 GPU, indicating a broader problem with the setup rather than a specific GPU issue.