Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.

MacVidCards

Suspended
Original poster
Nov 17, 2008
6,096
1,056
Hollywood, CA
Been debating where and when to post this.

Juicy.

Shows that Apple has a lab somewhere that is working on eGPUs. The very thing I posted about a few weeks ago. (CUDA coming to nMP near you)

I decided to try out a variety of cards as they are lying around here.

I can write an EFI that gives boot screen, etc but unlike cMP these cards still have no info in the "PCI" section of Profiler.

Instead, the usual "There was an error...blah, blah,blah" of unflashed cards.

Imagine my surprise when I tossed in an old Quadro 3800 and saw the attached screen. Why oh why would the "Graphics and Displays" section have those TB specific fields? Somebody put them there and it wasn't me.

I would guess that Apple has been quietly working on this for some time but has held off introducing either as a favor to Intel, to sell more computers, or just because they are mean.

In any case, the system profiler knows what an eGPU is. I would guess that they have a routine in that blocks this info on more common and recent cards, but somehow older obscure cards still show up like this.

We will be offering a limited number of eGPU options, but my guess is that eventually Apple themselves will. Or maybe not, I have found that my rMBP becomes an awesome gaming machine when I dock it with TB GTX980. Then I undock and take with me. At which point my 2014 dog-slow Mini becomes a gaming machine, via the same GTX980.

So, you could buy 1 pricey GPU and use it with all of your TB equipped Macs. When you wanted to run CUDA on your nMP, BANG ! Crisis 4 on your MBA? BANG ! Sounds far too useful and sensible for Apple to market it.

In any case, there it is, proof it is either coming or was tested and set aside.

We'll see which.
 

Attachments

  • Screen Shot 2014-11-23 at 2.32.12 AM.png
    Screen Shot 2014-11-23 at 2.32.12 AM.png
    73.6 KB · Views: 1,751
Very interesting...

Am I interpreting the screenshot correctly? It would appear that you added the efi rom modifications, but did not alter the driver info.plist to enable tunneling. Or did the profiler show this with no efi rom modifications at all?

I have an egpu setup using a R290X and would love to help this effort in any way possible. Let me know if I can help.

JimJ
 
Am I interpreting the screenshot correctly? It would appear that you added the efi rom modifications, but did not alter the driver info.plist to enable tunneling. Or did the profiler show this with no efi rom modifications at all?

I have an egpu setup using a R290X and would love to help this effort in any way possible. Let me know if I can help.

JimJ

What happens when you look in System Profiler under "PCI"?

I am only working on Nvidia cards and they all show same "Error..." in PCI section like an unflashed card, whether they have an eGPU EFI or not.

This Q3800 is unique in that not only does an unflashed card show up here, but shows Thunderbolt GPU specific info. I did the tunnel kext fix in order to have card show up at all, it could be that these could work better if we knew how to get those fields to say "yes'.
 
What happens when you look in System Profiler under "PCI"?

I am only working on Nvidia cards and they all show same "Error..." in PCI section like an unflashed card, whether they have an eGPU EFI or not.

This Q3800 is unique in that not only does an unflashed card show up here, but shows Thunderbolt GPU specific info. I did the tunnel kext fix in order to have card show up at all, it could be that these could work better if we knew how to get those fields to say "yes'.

I get an error message if I look in the PCI section, although the R290X is visible in the details of the ThunderBolt section.
 
I don't see anything unusual here for a Thunderbolt device. This is all what I'd expect to see without anything extra added for GPUs.

There are a lot of problems with Thunderbolt GPUs being officially supported by OS X. The GL subsystem just isn't compatible enough with hot swap GPUs, but it's getting better. Apps are getting better with swapping GPUs due to the Macbook Pros which hot swap GPUs, but you also have a huge problem in that you can't copy all the data out of VRAM to another GPU after the original GPU has been disconnected.
 
I don't see anything unusual here for a Thunderbolt device. This is all what I'd expect to see without anything extra added for GPUs.

There are a lot of problems with Thunderbolt GPUs being officially supported by OS X. The GL subsystem just isn't compatible enough with hot swap GPUs, but it's getting better. Apps are getting better with swapping GPUs due to the Macbook Pros which hot swap GPUs, but you also have a huge problem in that you can't copy all the data out of VRAM to another GPU after the original GPU has been disconnected.

I've never seen a screen shot like this for a GPU, maybe they are run of the mill.

In any case, I have never understood the fixation with "hot swappable".

If I have a TB HD or for that matter a USB HD, I know that I should "eject" it first. I also know that if I choose to yank the cord anyway I will be the one who caused the system to crash or my data to go Bye-bye. So why do GPUs need to adhere to "hot swap" while HDs connected same way don't ?

One of my primary issues with TB is the silly bunch of rules connected with it. The "there must be a video stream" being one of them and the "must be hot swappable" being another. If you look at an eGPU as being what it is, a GPU that has been connected to 4 lanes of PCIE interface and that you should only remove or disconnect it while the machine is off, all the problems go away.

Is this as good as plugging into 16 lanes of PCIE 2 or even 3? Absolutely not. But being able to run serious apps (or games) with a computer that normally talks to the world through integrated GPU is a revelation.
 
I've never seen a screen shot like this for a GPU, maybe they are run of the mill.

Thunderbolt will just forward whatever the PCI-E information is. I have a Thunderbolt array that shows up as a SCSI PCI card whenever I plug it in. So nothing really unusual here as far as I can tell.

In any case, I have never understood the fixation with "hot swappable".

If I have a TB HD or for that matter a USB HD, I know that I should "eject" it first. I also know that if I choose to yank the cord anyway I will be the one who caused the system to crash or my data to go Bye-bye. So why do GPUs need to adhere to "hot swap" while HDs connected same way don't ?

The problem is to swap a GPU requires a shut down of the machine, at least for OS X. I'm not sure how Windows does it, maybe it has to do with the advancements they made in Vista for being able to swap out GPU drivers on the fly. But it's a little worse than the case you'd have for a USB drive.

It's probably not a big deal for a lot of people, but for Apple, who'd love to be able to have people just yank a laptop from a GPU like they do with a Thunderbolt display, it will be a problem.

One of my primary issues with TB is the silly bunch of rules connected with it. The "there must be a video stream" being one of them and the "must be hot swappable" being another. If you look at an eGPU as being what it is, a GPU that has been connected to 4 lanes of PCIE interface and that you should only remove or disconnect it while the machine is off, all the problems go away.

Is this as good as plugging into 16 lanes of PCIE 2 or even 3? Absolutely not. But being able to run serious apps (or games) with a computer that normally talks to the world through integrated GPU is a revelation.

I don't disagree that even without 16 lanes, an eGPU on a Mac Mini is a huge step up and makes that system usable for a lot of users probably even here in the Mac Pro forums.

I do kind of understand the video requirement. Apple and Intel are pushing for smaller machines with more ports, and it makes sense to not have an HDMI port (or ports, cause c'mon, I want to hook two displays to my MBP) competing for space. Would I rather have Thunderbolt not have the video requirement, or be able to hook three displays to my Macbook Pro? I'd definitely prefer the three displays, with the flexibility to reuse those ports for data if I want.

My only serious complaint is that no one apparently thought about PCI-E GPUs when they came up with that whole plan. Intel should have really pushed for some sort of firmware support for re-channeling a video stream through the PCI-E bus back to the Thunderbolt bus. Even if it required new cards, Apple wouldn't have cared because they control the cards available any way. I was very surprised when I heard that the Mac Pro wouldn't have any PCI-E slots at all in the traditional PCI-E slot sort of way.

Ah well, at least Intel could still correct that in the future. I'm surprised they aren't actually pushing for an entirely internal PCI-E standard like what's on the nMP. Would be cleaner for everyone.
 
Not going to happen. Not from Apple anyway. They don't want you ruining their design ethos by adding an external PCIE box on your desk along side all the other external devices they forced you to adapt to. A mess of cable connected peripherals that makes PC owners gag.

There is however a dockable screen that Apple has patented and probably has in the labs. It's a screen with a slot for a slim notebook, turning something like a MacBook Air into an iMac. The eGPU would be built into the screen in this case. This might not get to market though.

080103_patent_app_monitor_dock.gif
 
Last edited:
Not going to happen. Not from Apple anyway. They don't want you ruining their design ethos by adding an external PCIE box on your desk along side all the other external devices they forced you to adapt to. A mess of cable connected peripherals that makes PC owners gag.

I wouldn't say Apple is likely opposed to it for design reasons (your snark aside.) After all, the design of the most recent Mac Pro is based around filling up your desk with boxes.

At this point it's entirely because the OpenGL subsystem in OS X isn't compatible with pluggable PCI-E cards, and Apple isn't going to sanction eGPUs until they can fix that. And fixing that depends on being willing to put in the amount of work to do it.

At the end of the day, I think their install base in laptops might drive them to do it. But as of now, even if Apple wanted to put a GPU into a Thunderbolt display, the OS just can't support it.
 
At the end of the day, I think their install base in laptops might drive them to do it. But as of now, even if Apple wanted to put a GPU into a Thunderbolt display, the OS just can't support it.

I added a relevant paragraph to my post.
 
Seriously, who draws these patent images? I mean, it's like they drew it on a moving train without even another pad of paper to use as a straightedge, or drunkenly scribbled out at a bar one night.
 
Even if there was an official eGPU from Apple, I'd imagine it like a (unibody-formed) closed, not-user-upgradeable metal box with a soldered, glued and nailed on top GPU inside.

I'd bet this will only be possible as a custom non-official solution.
 
Even if there was an official eGPU from Apple, I'd imagine it like a (unibody-formed) closed, not-user-upgradeable metal box with a soldered, glued and nailed on top GPU inside.

I'd bet this will only be possible as a custom non-official solution.

They will do their damndest to do that I think or it will be inside a 5k eGPU TB3 display based on the iMac 27" chassis
 
I added a relevant paragraph to my post.

I don't see that being a thing they ever do. That was before Thunderbolt. Why have that sort of setup now that you have Thunderbolt? It's the same thing as a Thunderbolt display, except with a clunky docking bay that will somehow have to be compatible with a bunch of laptop sizes.

You're more likely to see a GPU getting added to the Thunderbolt display, but again, the OpenGL stack in OS X is a problem. And Apple is having enough trouble dedicating resources to even keeping the OpenGL issue current.

It's not a problem of "Apple just wants to control the hardware so they aren't allowing third parties to do it."
 
Here's why

Chucked a couple Titan Blacks into the Mini.

OpenCl score matches new 12 Core with Dual D700s. (Only thing faster is a 2010 sporting some 7970s)

http://www.barefeats.com/tube08.html

At a savings of enough for a cheap week in Hawaii.

Obviously there are other factors, but if all you need is an OpenCl powerhouse, why pay $5K extra for a shiny buff job? (please no crude humor)
 

Attachments

  • Screen Shot 2014-12-05 at 1.32.28 AM.png
    Screen Shot 2014-12-05 at 1.32.28 AM.png
    172.4 KB · Views: 551
  • Screen Shot 2014-12-05 at 1.53.24 AM.png
    Screen Shot 2014-12-05 at 1.53.24 AM.png
    67.8 KB · Views: 403
  • Screen Shot 2014-12-05 at 1.27.21 AM.png
    Screen Shot 2014-12-05 at 1.27.21 AM.png
    1.1 MB · Views: 482
  • Screen Shot 2014-12-05 at 1.57.02 AM.png
    Screen Shot 2014-12-05 at 1.57.02 AM.png
    99.7 KB · Views: 380
Thanks MVC. That cheapo six core tower destroys the ridiculously expensive trashcan. A single GTX 980 scores the same as the Dual D700 if paired with similar CPU power.
 
Last edited:
Seriously, who draws these patent images? I mean, it's like they drew it on a moving train without even another pad of paper to use as a straightedge, or drunkenly scribbled out at a bar one night.

I can't even draw like that. I wish I could draw that neatly. :D

----------

What TB Enclosure are you using for your eGPU?

I am also curious about this.
 
The 10 series successor to the GTX 980 will be even more powerful more efficient and of course compatible with Mac Pro towers. It's the true successor to the seven series. The nine series was just a modest transition to efficient design.

Then there is Pascal. There might be a version compatible with PCIE 1/2/3 but it's true performance needs much more bandwidth. Currently there aren't apps or games or storage devices that can feed so much data so quickly anyway.
 
Last edited:
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.