eGPU on MacBook Pro 4,1 and up with PCIe ExpressCard slot

[Larsvonhier]

Hello dear Larsvonhier,

for many years I am upgrading my machine (MacBookPro5,1) some weeks before I was looking into if I could upgrade the processing power of my laptop and stumbled upon your videos (

) where you were showcasing that you successfully got an eGPU to work with your MacBookPro4,1. I got inspired on setting things up by your example via the ExpressCard34-Slot. Since some weeks now I am trying to get everything to work properly but I still keep failing at it. Then I found your thread here and didn't want to hesitate to maybe ask you for your help (since you even got everything worked out on an older MBP).

At this moment I got macOS Catalina 10.15.7 (19H2) working on my MacBookPro5,1 from Late 2008 by using DosDude1's Patcher. Everything is running smoothly except for some bugs here and there which go away after some time (mainly only coming up at the start of the system). I already got the EXP GDC Beast version 8.5c where I want to that I wanted to hook up with the Gainward Phantom (GeForce) GTX 770 working as eGPU. The GraphicCard seems to be natively compatible with Catalina because of the KeplerChip, but the machine doesn't recognize it - although it's juiced up via the ATX-Power module. The fans are also running on full speed. I installed the latest CUDA Driver from NVidia but there are no WebDrivers from NVidia that would work. When trying to install them it only says "Mac OS X version is not compatible // Mac OS X version 10.15.7 (19H2) is not supported with this package. Please see NVIDIA’s website for further driver information."

Then I read that the NVidia drivers are only supported up to High Sierra which led me to try and install a dual boot with Catalina for my main tasks and High Sierra for cutting video and other intensive work. But that didn't workout fine. It seems like the HighSierra Patcher from DosDude1 only allows one EFI and so my Catalina EFI got deleted and I only could get into HighSierra. But even there I couldn't get the eGPU to work/being recognized by the system. So I had to reinstall the functioning Catalina system back from my TimeMachineBackups.

Now you mentioned that you and/or your colleague from work still got DosDude1's Catalina to work fine with an eGPU Nvidia GraCa and I wanted to ask if you could help me with this?

Is there any solution? One thing I found is Lilu and the WhateverGreen Kexts which seem to need a Clover based EFI and as far as I know DosDude1's Patcher does not utilize Clover, right? But then I read that one could implement Boot Arguments from the Terminal and so the CloverConfigurator would become obsolete. Do you have some suggestions? Or do you think it would be a better choice to go with an AMD card? And if so which would you recommend?

Really appreciating your dedication, work and help to keep old hardware alive and running macOS with more graphical power.
Sorry if my English is not that good. I'm from Germany, too, and tried to keep it helpful for others that might search the web for similar answers

Best Regards,
Arthur

Hi Arthur,

sorry for delayed response, there´s too much work load on my work days right now...
Good news: Yes, there is a way to get 10.15.6 (and .7) up and running with eGPUs, with expresscard34 solutions like the GDC beast and also (on newer machines) via thunderbolt-1 (!).
As you asked for the expresscard way to do it: The "secret" is to use Nvidia cards that are based on chipsets that are same or close to what Apple included in their earlier (2008..2010) laptops -> 8800M/GT, 9400M. These cards will then use OpenGL after patching for it with dosdude1´s tools. There is currently no way to patch the internal GPU for OpenGL and let the eGPU run with metal - but hey, there is no real disadvantage as the eGPUs running OpenGL are still way faster than their thermally throttled internal cousins. The only way to avoid this setback is to stick with 10.13.x and the available Nvidia drivers there. (It´s a shame that Apple and Nvidia behave like that and play kindergarden).

If you´re looking for the thunderbolt solution, I´d have to compile some step-by-step instructions and infos on the tools used (not sure if I used the then-current versions).

Let me know if this helped or where you´re stuck (and with what GPU)...

 

[Camillione]

Hi Arthur,

sorry for delayed response, there´s too much work load on my work days right now...
Good news: Yes, there is a way to get 10.15.6 (and .7) up and running with eGPUs, with expresscard34 solutions like the GDC beast and also (on newer machines) via thunderbolt-1 (!).
As you asked for the expresscard way to do it: The "secret" is to use Nvidia cards that are based on chipsets that are same or close to what Apple included in their earlier (2008..2010) laptops -> 8800M/GT, 9400M. These cards will then use OpenGL after patching for it with dosdude1´s tools. There is currently no way to patch the internal GPU for OpenGL and let the eGPU run with metal - but hey, there is no real disadvantage as the eGPUs running OpenGL are still way faster than their thermally throttled internal cousins. The only way to avoid this setback is to stick with 10.13.x and the available Nvidia drivers there. (It´s a shame that Apple and Nvidia behave like that and play kindergarden).

If you´re looking for the thunderbolt solution, I´d have to compile some step-by-step instructions and infos on the tools used (not sure if I used the then-current versions).

Let me know if this helped or where you´re stuck (and with what GPU)...

Hello Lars,

thanks for your reply.
That's nice to hear that it still would be possible to get an eGPU running on my system.
So after you mentioned OpenGL, I researched a little bit and found out that I got Version 3.3 and Version 1.2 for OpenCL running. I'm not sure if it's possible to upgrade OpenGL/CL, but read that it is being supported up to 4.1 by macOS. So after that I guess GraCa's that support v3.3 (GL) and v1.2 (CL) should work fine. Then I took a look at the GPU-Processors and saw that the ones that run fine with the versions are GT218 and GF108 (https://de.wikipedia.org/wiki/Liste_der_Nvidia-Grafikprozessoren). Then I tried to find out to what Chip family they are being connected within the nVidia universe and they seem to be part of the Tesla 2.0 and Fermi family (https://www.techpowerup.com/gpu-specs/?architecture=Tesla 2.0&sort=generation). Scrolling through the list I stumbled upon GraCa's that included the term "Mac Edition" so they seemed to be specifically made for Apple's Macs. Looking one up I noticed that even better GPU processors should see support by Catalina as they should get native kexts and so on (https://www.techpowerup.com/gpu-specs/geforce-gtx-680-mac-edition.c2323). Looking at CUDA support via nVidia it is being capped by High Sierra. So that route can be neglected I think.
So either one should concentrate on limits the installed version of OpenGL provides (is it even upgradable? or will higher versions be locked open by using a better GraCa?) or should I better consider to take a look at AMD as they seem to be supported natively by macOS Catalina? Activating Metal by this the constrictions described above would not being applied anymore right? But then I am asking myself if the system will even recognize the GraCa via the EXC GDC Beast. I just don't want to buy GraCa's that I think of should work out fine but then will fail because of further compatibility issues.
What would be your guess? What should I be fixing on? OpenGL version + nVidia Card or AMD + Metal? And is Metal even activated inside of DosDude's patched Catalina?

PS: Or would the screenshot of the last supported GraCa's supported via CUDA 10.2 be a better hint (https://de.wikipedia.org/wiki/CUDA )? But then I don't understand why the GTX 770 won't work except for that it just got support up to HS. Also I saw some that got CUDA and the WebDrivers somehow working by NVRAM boot flags and so on, but I would prefer to just go the native route so that I don't brick my machine.

Best Regards
Arthur

 

[Larsvonhier]

Hello Lars,

thanks for your reply.
That's nice to hear that it still would be possible to get an eGPU running on my system.
So after you mentioned OpenGL, I researched a little bit and found out that I got Version 3.3 and Version 1.2 for OpenCL running. I'm not sure if it's possible to upgrade OpenGL/CL, but read that it is being supported up to 4.1 by macOS. So after that I guess GraCa's that support v3.3 (GL) and v1.2 (CL) should work fine. Then I took a look at the GPU-Processors and saw that the ones that run fine with the versions are GT218 and GF108 (https://de.wikipedia.org/wiki/Liste_der_Nvidia-Grafikprozessoren). Then I tried to find out to what Chip family they are being connected within the nVidia universe and they seem to be part of the Tesla 2.0 and Fermi family (https://www.techpowerup.com/gpu-specs/?architecture=Tesla 2.0&sort=generation). Scrolling through the list I stumbled upon GraCa's that included the term "Mac Edition" so they seemed to be specifically made for Apple's Macs. Looking one up I noticed that even better GPU processors should see support by Catalina as they should get native kexts and so on (https://www.techpowerup.com/gpu-specs/geforce-gtx-680-mac-edition.c2323). Looking at CUDA support via nVidia it is being capped by High Sierra. So that route can be neglected I think.
So either one should concentrate on limits the installed version of OpenGL provides (is it even upgradable? or will higher versions be locked open by using a better GraCa?) or should I better consider to take a look at AMD as they seem to be supported natively by macOS Catalina? Activating Metal by this the constrictions described above would not being applied anymore right? But then I am asking myself if the system will even recognize the GraCa via the EXC GDC Beast. I just don't want to buy GraCa's that I think of should work out fine but then will fail because of further compatibility issues.
What would be your guess? What should I be fixing on? OpenGL version + nVidia Card or AMD + Metal? And is Metal even activated inside of DosDude's patched Catalina?

PS: Or would the screenshot of the last supported GraCa's supported via CUDA 10.2 be a better hint (https://de.wikipedia.org/wiki/CUDA )? But then I don't understand why the GTX 770 won't work except for that it just got support up to HS. Also I saw some that got CUDA and the WebDrivers somehow working by NVRAM boot flags and so on, but I would prefer to just go the native route so that I don't brick my machine.

Best Regards
Arthur

Some things to be sorted out for better understanding (sorry, no or short explanations as I´m a bit in a hurry today):
- When you patch with dosdude1´s patcher (for Mojave or Catalina) and enable "legacy graphics" patch, you will not be able to use Metal on any card (internal or eGPU) any more, as underlying frameworks are reverted back to OpenGL
- OpenCL and CUDA is mostly irrelevant for system performance as long as you don´t do GPU based calculations (number crunching, mining, etc.)
- Mac Edition cards only differ from their PC counterparts in that they have a special boot-flash that enables the boot screen for Macs. Especially in eGPU usage (where you normally have the internal card and macbook screen up and running at boot time) there is no need to get the more expensive and rare Mac edition cards
- by adding NVRAM parameters, boot args or other entries you can not permanently brick your machine - there is always the way back by doing a NVRAM reset at cold start of your machine (command-P-R keys)

Beste Grüße,
Lars

 

[Camillione]

Some things to be sorted out for better understanding (sorry, no or short explanations as I´m a bit in a hurry today):
- When you patch with dosdude1´s patcher (for Mojave or Catalina) and enable "legacy graphics" patch, you will not be able to use Metal on any card (internal or eGPU) any more, as underlying frameworks are reverted back to OpenGL
- OpenCL and CUDA is mostly irrelevant for system performance as long as you don´t do GPU based calculations (number crunching, mining, etc.)
- Mac Edition cards only differ from their PC counterparts in that they have a special boot-flash that enables the boot screen for Macs. Especially in eGPU usage (where you normally have the internal card and macbook screen up and running at boot time) there is no need to get the more expensive and rare Mac edition cards
- by adding NVRAM parameters, boot args or other entries you can not permanently brick your machine - there is always the way back by doing a NVRAM reset at cold start of your machine (command-P-R keys)

Beste Grüße,
Lars

Hi Lars,
thanks for the fast reply and your clarification.
Now I get it. So depending on Catalina with dosdude patch I can only accelerate graphics only via OpenGL (and it looks like only up to version 3.3/4.1); so I would have to focus on Chips and Cards that support these underlying constraints.
Would this stay in effect with Catalina on OpenCore with Lilu & Whatevergreen Kexts? To my understanding Metal would then still be active, right?
Another possible way would be to install High Sierra on a secure drive and do the video work from there with CUDA+WebDrivers. Yet dosdude's simple patcher would not fit for that purpose I guess (as I already tried and could get back into Catalina; I'm not that savy in setting up different EFI's and parallel OS). But maybe an OpenCore HS installation would be the right fit here, too.
Another thought that came to me were VirtualMachines and getting HS or Catalina to work with Metal/ CUDA on there, but my trial with HS didn't lead to a positive outcome either, as the eGPU wasn't recognized by the main OS in the first place (Catalina).
Could Crossover + the Windows Version of CUDA maybe be sufficient enough to get it working on macOS? But I don't think that this could be that easy of a workaround.
Beste Grüße
Arthur

 

[TouchMeSenpai]

Hello everyone! I have an early 2011 15" i7 Macbook Pro (8,2) and after reading through this thread I've decided to purchase an exp gdc myself, but I would like to run my plans by here first and ask for advice to minimize any mistakes I might make in purchasing and save time in troubleshooting as well

-I'm planning on purchasing this version as this seems to be the one that @kautame has had the least problems with if I read correctly
-I'll be sacrificing my wifi card and inserting the exp gdc through the mPCIE slot
-A gtx 750 ti is what I'm planning to use as it's the easiest card to find in my area within my budget
-I have a 300w PSU from a previous home computer my family used

Is this okay? Is it feasible? Are there any parts that won't work?


A few questions I have:

*I noticed that most of the questions about installation asked here are about installing it on mac os, however I'll mostly be using Windows 10 as I intend to play a few games other than my main purpose of video editing, I can't find an exact precursor to what I'm planning to do but I found this and this, are these the guides that I should follow? Will the process be as straightforward for me as in the first link event though I'm inserting it through the mPCIe instead of ExpressCard? Any other points I should watch out for?

*Do you have other GPU suggestions? Anything close to the 750 TI in performance is good enough for me as I really only play Monster Hunter: World, so anything easy or has been tested to work the best/easiest would be great.

Thank you so much for taking the time to read this and hopefully reply. Have a great day!

 

[Corvinon]

Some things to be sorted out for better understanding (sorry, no or short explanations as I´m a bit in a hurry today):
- When you patch with dosdude1´s patcher (for Mojave or Catalina) and enable "legacy graphics" patch, you will not be able to use Metal on any card (internal or eGPU) any more, as underlying frameworks are reverted back to OpenGL
- OpenCL and CUDA is mostly irrelevant for system performance as long as you don´t do GPU based calculations (number crunching, mining, etc.)
- Mac Edition cards only differ from their PC counterparts in that they have a special boot-flash that enables the boot screen for Macs. Especially in eGPU usage (where you normally have the internal card and macbook screen up and running at boot time) there is no need to get the more expensive and rare Mac edition cards
- by adding NVRAM parameters, boot args or other entries you can not permanently brick your machine - there is always the way back by doing a NVRAM reset at cold start of your machine (command-P-R keys)

Beste Grüße,
Lars

Hi Larsvonhier
So if I put a gtx660m (PC-Kepler) on the imac, I just lose the image during boot?
And another question, my mxm is dead, if I put a gt710 via pci (wifi card) it will boot directly from it, not being able to do it from the mxm?
A million thanks

 

[Larsvonhier]

Hm...
answer to your first question: That should be correct - it´s just a major nuisance from any point on (updating/upgrading macOS) when you have to repatch or troubleshoot things. One relief here is that the patcher sticks created by dosdude1´s tools are capable to turn on graphics on non-bootscreen cards during late boot phase, so you can at least re-install and patch things from there.

Don´t really understand your second question... But I´ll try to interpret:
If you install an eGPU (GDC beast) through the miniPCIe port (normally hosting the Wifi/Airport card), and get the system to boot normally (despite broken/dead MXM slot GPU), with the right software/driver ingredients it should be usable also. What might go wrong is a semi-dead MXM card leading to instabilities/crashes or lockups during boot. Might be even better to take it out completely. But I can only guess here...

 

[Corvinon]

Hm...
answer to your first question: That should be correct - it´s just a major nuisance from any point on (updating/upgrading macOS) when you have to repatch or troubleshoot things. One relief here is that the patcher sticks created by dosdude1´s tools are capable to turn on graphics on non-bootscreen cards during late boot phase, so you can at least re-install and patch things from there.

Don´t really understand your second question... But I´ll try to interpret:
If you install an eGPU (GDC beast) through the miniPCIe port (normally hosting the Wifi/Airport card), and get the system to boot normally (despite broken/dead MXM slot GPU), with the right software/driver ingredients it should be usable also. What might go wrong is a semi-dead MXM card leading to instabilities/crashes or lockups during boot. Might be even better to take it out completely. But I can only guess here...

I said, thank you very much.
Your answer has helped me a lot.

 

[gonvelho]

Hi, blessed by this entry I bought all the material.
- Edx gdc Beast 8.5c
- GPU GeForce GTX 650
Running on Macbook pro 17'' early 2011 (with deactivated dgpu and running previous just the internal).
In High Sierra it was first not detected. After installing drivers, detected but no external monitor image (with VGA DVI cable).
Installed dosdude patched version of Mac OS Catalina. Egpu card is detected, but still no external image with VGA-DVI cable. Apps run well, namely the ones that need Metal. EGPU is recognised in GFX Metal, but no so good results.
All advices are welcomed.
Main aim was using it with Prezi Video and Zoom and it works like a charm with a finnaly upgraded version of MacOS to Catalina.
Thanks to everyone contribution.

 

[Larsvonhier]

Try digital output like DVI (without VGA adapter) or DP/HDMI. Some card/driver/patch combinations output only to certain ports, not all of them. Make sure that when patching with dosdude1´s package you do not select "legacy video patch", otherwise certain frameworks will be changed from metal to OpenGL. Result (if done otherwise) is metal capability of your machine, but only for apps "running" on internal screen (via OpenGL, though!). No output then through the eGPU.
As there is no easy way back once patched for legacy gfx, I'd suggest a fresh installation of Mojave or Catalina, patch without legacy gfx and see then ;-)
(Your internal screen will then be non-accelerated, but all eGPU connected screens will be!).
Unfortunately, you can't have best of both worlds.

 

[gonvelho]

Try digital output like DVI (without VGA adapter) or DP/HDMI. Some card/driver/patch combinations output only to certain ports, not all of them. Make sure that when patching with dosdude1´s package you do not select "legacy video patch", otherwise certain frameworks will be changed from metal to OpenGL. Result (if done otherwise) is metal capability of your machine, but only for apps "running" on internal screen (via OpenGL, though!). No output then through the eGPU.
As there is no easy way back once patched for legacy gfx, I'd suggest a fresh installation of Mojave or Catalina, patch without legacy gfx and see then ;-)
(Your internal screen will then be non-accelerated, but all eGPU connected screens will be!).
Unfortunately, you can't have best of both worlds.

Thanks Yep, that was it: when patching with dosdude1 package I didn't do the legacy video patch?‍♂️ No problem, since for now my aim was to have metal on this screen and work with the above mentioned apps (and they work OK, system is fluid, all running smooth).
Anyway, tried the DP/HDMI just to be sure and confirmed: no luck for external screen with legacy video patch.
As I ordered a box for installing a spare SSD on the DVD slot, I will then try to install in a second drive the correctly patched Catalina. Then I can move between boots, if either I want internal screen, or external.
It's incredible how with 60€ got this 10y old MBP running like a charm. Thumbs up!

 

[gonvelho]

Ok, I installed a new version clean. While installing the image goes to second monitor, but as soon as it boots from the installed version it's gone.
I tried to make the post install version without the legacy video, but still no luck...
Any clues?

 

[Larsvonhier]

Ok, I installed a new version clean. While installing the image goes to second monitor, but as soon as it boots from the installed version it's gone.
I tried to make the post install version without the legacy video, but still no luck...
Any clues?

Do you have the possibility to check back your configuration with High Sierra? That should work oob and if so,
it might be a driver issue under Mojave/Catalina (Nvidia web drivers not available but possibly needed).
Other GFX cards that do not need such drivers (or, drivers included in macOS) would be the solution then.

 

[gonvelho]

I installed previously the Nvidia drivers on the High Sierra that is running with internal screen. It shows that's unavailable the web drivers and is running only with MacOS. As my intention is to use a certain software that requires Metal and not so much performance, I'm happy for the current situation, even if without external monitor.
Probably, on as a saturday morning entertainment I might try the High Sierra install on another disk If that happens I will give feedback.

 

[Krutav]

Hello, really liked the outcome of being able to plug in a GPU and have it work. Personally, I wanted to do this project with a custom adapter for the WiFi port. (I will design a PCB with PCIe switching to retain the WiFi capability)

My question is, can I use the eGPU as a graphics accelerator for the internal screen? I have the 2011 MBP, so I am using the internal LCD with Intel HD 3000, and the eGPU will be AMD GPU. How would I use the eGPU to accelerate apps like metal apps and games without needing an external monitor? (It's small and compact, no space for a monitor!)

 

[Amethyst1]

My question is, can I use the eGPU as a graphics accelerator for the internal screen?

How To: eGPU-Accelerated Internal Display in macOS | eGPU.io

egpu.io

 

[Krutav]

How To: eGPU-Accelerated Internal Display in macOS | eGPU.io

egpu.io

Thanks for the suggestion. I considered this before, but I actually want to sell this MacBook with the GPU in it later on, so that it works out of the box. Needing the dummy dongle display and duplicating the monitor doesn't seem like the cleanest solution when the Thunderbolt eGPU docks can be used as a 3D controller out of the box, like how Apple shows in their documentation.

My goal is to just use the GPU as an accelerator, which means the Intel HD 3000 will drive the display, and any Metal apps or games get rendered on the GPU and use Apple framebuffer sharing. (works with thunderbolt, and works with any GPU on Windows 10)

 

[Larsvonhier]

Thanks for the suggestion. I considered this before, but I actually want to sell this MacBook with the GPU in it later on, so that it works out of the box. Needing the dummy dongle display and duplicating the monitor doesn't seem like the cleanest solution when the Thunderbolt eGPU docks can be used as a 3D controller out of the box, like how Apple shows in their documentation.

My goal is to just use the GPU as an accelerator, which means the Intel HD 3000 will drive the display, and any Metal apps or games get rendered on the GPU and use Apple framebuffer sharing. (works with thunderbolt, and works with any GPU on Windows 10)

Strongly depends on what macOS you intend to use it with. I once had High Sierra running in such a setup (internal screen using OpenGL API but metal capable eGPU for acceleration). With macOS that needs patchers like i.e. dosdude1´s you'll probably end up with totally unaccelerated internal screen but computing-accelerated for metal apps.
Also note: Some apps can be set to use eGPU instead of internal one, even when running on the internal screen.

As for sharing the Wifi miniPCIe slot for the eGPU and the Wifi card: It's only one PCIe lane so that might only work if a USB connection for Wifi is added (some adaptors for continuity-capable cards work that way).

 

[Krutav]

As for sharing the Wifi miniPCIe slot for the eGPU and the Wifi card: It's only one PCIe lane so that might only work if a USB connection for Wifi is added (some adaptors for continuity-capable cards work that way).

I'll be using a PCIe switch, so it will all work out fine.

Also note: Some apps can be set to use eGPU instead of internal one, even when running on the internal screen.

So I read that if you use a real Thunderbolt eGPU, there will be a button on the menubar that shows the GPU. It will also allow you to set that eGPU for all apps. It can also accelerate the internal display. But does it show with a PCIe eGPU?

 

[Larsvonhier]

I'll be using a PCIe switch, so it will all work out fine.



So I read that if you use a real Thunderbolt eGPU, there will be a button on the menubar that shows the GPU. It will also allow you to set that eGPU for all apps. It can also accelerate the internal display. But does it show with a PCIe eGPU?

The menu bar icon is there for the PCIe-attached GPU also. But you can't really do anything with it. Just informative (eject card option might reset the system).
Instead, each app can be set to use the eGPU individually (for apps that support that) by showing info about the app in finder. There's a selector checkmark box called "always use external GPU" or similarly.

Out of interest: How does the PCIe switch work? Can it attach two targets to one PCIe lane simultaneously?

 

[Krutav]

Instead, each app can be set to use the eGPU individually (for apps that support that) by showing info about the app in finder. There's a selector checkmark box called "always use external GPU" or similarly.

Wow thank you so much! I was not expecting the menu bar item to show with PCIe eGPU, but knowing it does solves the problem partially.

Out of interest: How does the PCIe switch work? Can it attach two targets to one PCIe lane simultaneously?

PCIe switches are relatively common as they are used when you want a lot of PCIe Lanes/devices with a limited number of lanes at the source. For example, you have 1 lane of PCIe, but have 2 devices, the PCIe switch will present both those devices to the system just like a network ethernet switch.

 

[Larsvonhier]

Wow thank you so much! I was not expecting the menu bar item to show with PCIe eGPU, but knowing it does solves the problem partially.



PCIe switches are relatively common as they are used when you want a lot of PCIe Lanes/devices with a limited number of lanes at the source. For example, you have 1 lane of PCIe, but have 2 devices, the PCIe switch will present both those devices to the system just like a network ethernet switch.

Ah, right. Thanks! I found a couple of ICs that do that trick - by limiting/sharing the bandwidth of that one lane...
(sounds good for I/O stuff, but not that great for eGPUs where one full lane is already a bottleneck).

But very interesting what the final PCB will look like that you want to design. Keep us in the loop!

 

[Krutav]

But very interesting what the final PCB will look like that you want to design. Keep us in the loop!

You're right about the huge bandwidth bottleneck. One PCIe Gen 2 lane is capable of 500MB/s bidirectional. If the wireless card uses 56MB out of the 500, we have 444 MB of bandwidth available for the GPU. This is pretty poor, but to be honest with you I am using a potato GPU (FirePro M4000) and I only need acceleration, no gaming work.

With that said, PCIe switches are ridiculously expensive, and my plan instead is to just use the PCIe lane from the FireWire port because nobody uses that now.

I have already a PCB design that is Open-Source for all of you and allows you to adapt an MXM laptop form factor GPU to a PCIe slot. I will reuse this design for the MacBook mod most likely. Here is the project on GitHub.