I'm thinking of switching to mac finally when the new Mac Pro comes out. I'm a hardware head, and always built my own PCs so it's going to be tough for me to let go of the personalized / superior hardware. But OSX wins I guess. Anyhow does anyone know anything about using ThunderBolt 2, to host an external GPU enclosure, so I can run a SLI setup for bootcamp? I don't think it would be a problem, but would I have to swap my monitors back to the Mac when booting OSX, or could I always run my displays on the external devices, and just use the internal GPUs for CUDA stuff? Also, the new Mac looks alot like a Dyson vacuum someone lifted from a janitor on the Death Star.
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
New Mac Pro / Thunderbolt / External GPUs
- Thread starter skyler.truax
- Start date
- Sort by reaction score
First off, the internal GPUs will almost certainly be AMD, which precludes using them for CUDA (an Nvidia API).
I'm no expert on GPUs over TB, but I can tell you that Thunderbolt (1 and 2) provide for a PCIe x4 connection, so even if GPUs are supported (I'm personally not sure) I doubt they will perform to their full potential.
Given my lack of helpful information, you might be wondering why I'm even replying... The truth is, I'm curious why you need SLI for bootcamp?... What's wrong with the installed AMD GPUs (which might support Crossfire in Windows)?... And is this for gaming? If so, is the new Mac Pro a wise choice as a gaming rig? You might be better of with a gaming PC that allows you to upgrade GPUs easily and then a Mac more suited to your other computing needs.
Thunderbolt 2.0 = 20 Gigabits per second
20 / 8 = 2.5 Gigabytes per second
PCIE 3.0 x16 = 15.7 Gigabytes per second
Thunderbolt 2.0 does not even provide PCIE 3.0 x4 (2.5 GBs vs 3.9 GBs)
So yeah, if you are a gearhead stay as far away from the new mac pro as you can. Seriously, its hardly a »Pro«*machine. Just a faster Mac Mini.
20 / 8 = 2.5 Gigabytes per second
PCIE 3.0 x16 = 15.7 Gigabytes per second
Thunderbolt 2.0 does not even provide PCIE 3.0 x4 (2.5 GBs vs 3.9 GBs)
So yeah, if you are a gearhead stay as far away from the new mac pro as you can. Seriously, its hardly a »Pro«*machine. Just a faster Mac Mini.
Well, I'm a gear head, but also a coder and developer. Truthfully I work on my computer a lot more than I play, so all things considered that should be the final decision. What it comes down to is OSX is a superior OS for a developer, and well for everyone. But I am very used to building my own very powerful, very custom machines. I have a triple monitor setup, and for certain games (WOW in particular) I use them as one device through Nvidias drivers. I know Crossfire / eyefinity does the same thing. I might not be opposed to just using the internal AMD devices but I was just seeking out options.
BTW I wasn't just throwing punches in the dark, people are experimenting with this:
http://hackaday.com/2013/07/31/a-macbook-air-and-a-thunderbolt-gpu/
And there is a company planning dual and even quad GPU external enclosures for thunderbolt 2....
http://www.sonnettech.com/product/thunderbolt/
I'm certain running it for windows wont be a problem. And while the Mac Pro is meant to be a workstation, I'm sure the 12 core Xenon will do just fine for the games I play If I can also run a good video rig. My main question is, will I be able to leave the monitors hooked up to the external cards when in OSX or will I have to swap the cables all the freaking time.
BTW I wasn't just throwing punches in the dark, people are experimenting with this:
http://hackaday.com/2013/07/31/a-macbook-air-and-a-thunderbolt-gpu/
And there is a company planning dual and even quad GPU external enclosures for thunderbolt 2....
http://www.sonnettech.com/product/thunderbolt/
I'm certain running it for windows wont be a problem. And while the Mac Pro is meant to be a workstation, I'm sure the 12 core Xenon will do just fine for the games I play If I can also run a good video rig. My main question is, will I be able to leave the monitors hooked up to the external cards when in OSX or will I have to swap the cables all the freaking time.
the 12 core Mac Pro is likely up in the $5-6K range. That is huge chunk capital costs to play games on. Frankly, buy at 6-8 core box and using the"left over" $1-2K to buy a gaming box would be a relatively sound approach.
Swap? There are enough Thunderbolt (TB) ports you could just simply pull the cable if don't want the hardware to appear while booted up in OS X.
You probably would want to reserve one of the TB controllers ( and the associated two ports) solely for your external PCIe card enclsoure (for GPU ). An isolated TB networked will give the enclosure full bandwidth. That would still leave you with 4 TB ports for whatever else. Including two for DisplayPort/DVI ports out. Get a video switch ( or use the video switch in the monitor) and don't have to swap anything.
Leaving your emotionally charged statement about the new Mac Pro out of this...
First, while Thunderbolt 2.0 does signal at 20Gbps, it only carries a PCIe 2.0 x4 data connection. So only 16Gbps of that is used for data... not the full 20Gbps.
Secondly, you imply that a modern GPU can utilize the full bandwidth of a 16-lane 3.0 bus... which of course they can only do in the most extreme compute oriented situations. Many have run GPUs externally via TB with some benefit as linked in the OP's post. Of course, you may not get full performance, but it does work fine in some applications.
So Thunderbolt is neither as capable as you suggest nor will the impact be as severe as you imply.
I'm sorry, but I'm not an expert on external GPUs over TB... so hopefully some others with more knowledge will chime in. It's possible you're even more knowledgeable on the subject than anyone else around here
It will be interesting to see how external GPUs over TB work out. However, I wouldn't discount the capability of the two included GPUs. They may provide all the GPU power you need.
Right but Thunderbolt 2 might change that a lot.
I'm probably going to wait til Mac Pro is out for 2 months and people have already tested what I'm envisioning...
I just wanted to see if anyone else was planning such a beast.
As for Price... I know the Mac Pro is not the best way to get the most hardware out of my dollar. I'm not really that concerned with cost. But I don't want 2 separate main workstations.
Pragmatically a bit less than that.
" ... Thunderbolt 2 should raise that to around 1500MB/s (overhead and PCIe limits will stop you from getting anywhere near the max spec). ... "
http://www.anandtech.com/show/7049/intel-thunderbolt-2-everything-you-need-to-know
around x3 PCIe v2.0. Although, if isolate a single controller to a single TB device which internally only uses x4 lane mode that would reduce that the overhead to absolute minimums. That is a straight-shot x4 to x4 pathway with nothing else on the wires.
Have there been any SLI/Crossfire external link up? Single cards yes, but two cards also increases the bandwidth limitation pressure.
In mainstream PCs working with just physical budget of x16 lanes typically the SLI/Crossfire set ups go x8 & x8 , or x4 & x4 & 8 , or x4 & x4 & x4. This set up is conceptually x2 & x2 and pragmatically pretty close if no careful to x1.5 & x1.5
The real weakly motivated question is why the external GPU. There are two data channels in Thunderbolt. One for PCIe and one for DisplayPort. Focusing only on one or the other will miss half the value proposition in using it. The DisplayPort stream is that in part because the design aligned usage is that the GPU isn't external. The output from the GPU is made external. So the who brouhaha over "but 2 x16 PCIe v3 bundles are faster" is moot because Mac Pro has and leverages those just fine.
So far it is far more like they "happen to work". Don't unplug anything and make it look like a static PCIe connect and will fool the Windows drivers enough to get a card (and display output) working. Use along the full spectrum of Thunderbolt utilization and it doesn't work.
As an alternative hack for limited Intel HD3000/4000 GPUs over time there is some utility. For not so limited fully clocked desktop GPUs over time there is alot less utility. It is a problem that Thunderbolt was never intended to solve.
Not really. The vast majority of Thunderbolt 2 deployments are going to be in context of host systems with one single controllers. What going to get in most deployed contexts is a mix usage of the Thunderbolt network stream. Either multiple Thunderbolt devices on same controller( in which overhead and bandwidth sharing issues will pop up) or DisplayPort usage which forces all the TB devices onto a single chain.
Far more likely folks are going to shift lower bandwidth hogs into external PCIe enclosures. Legacy Audio/Video capture/compression/transform cards.
In cost and physical space occupied not going to be a huge difference. The large-enough-for-two-midrange-GPUs enclosure isn't going to be small footprint. That Sonnet III box is $949. You can get a desktop PC for $949 as the foundation for two GPU cards you want.
A horizontal desktop PC with a Mac Pro resting on top still has the footprint of a horizontal desktop PC. Either way there is a second box on the desk.
It would sure be nice if there was an nvidia option in the nMP. We know Firepro is entering the OSX ecosystem, we don't know there won't be a Quadro or Geforce option also at the mid or base price points. Perhaps the Nov 15 date, corresponds with the R9 Volcanic islands release (AMD) .. one can dream.
I believe Nvidia has confirmed they are not in the new Mac Pro...
https://forums.macrumors.com/threads/1614432/
Why the concern about CUDA? If it's for use with Adobe applications they've recently confirmed that CC will be updated to use OpenCL on the new GPUs inside the Mac Pro. Adobe products should scream on the new Mac Pro without fail. Or is there another reason you're concerned about CUDA and not OpenCL?
Adobe products are the least of my concerns (I guess AFX using openCl will be nice). I'm thinking about NukeX, TurbulenceFD, Octane and various other rendering and simulation software that have CUDA acceleration. The fact is anyone offloading significant work onto the GPU is today using CUDA. That code is far more mature. Sure I can hope company X decides to also support openCL, but will they, and in what sort of time frame? How many years or months from now will those Firepro's have nothing to chew on? How good will those initial implementations be? Chances are I also need to buy a new version of the software that today's version already supports cuda. But looking at my software today I would put 2 Titan's into the nMP in a heartbeat.
Honestly I think OpenCL is great as a more open API, and I hope it gets adopted by a lot of developers. Maybe when the 2015 Mac pro comes out the landscape will be more openCl friendly.
I have a MP 1,1, I remember the big hoopla about 64 bit computing (started with the G5 really).. how many years did it take for the 64 bit applications (other than chess
) to actually show up?
Last edited:
I'm not into this space, but I'm curious... What do you think will happen? I assume these companies will either continue to offer Mac versions and add OpenCL support, or EOL the Mac versions and go strictly with Windows/CUDA because very soon, a Mac version of these kinds of programs with out OpenCL support makes no sense.
We'll have to wait and see if Nvidia is totally shut out of the nMP. If people hook up external GPU's etc. For now there are the 2012 MP's and earlier. The smart developers are trying to unify their code base. So they don't have 3 versions of their software to maintain. So they write the code once and deploy to multiple platforms. Long term openCL makes sense since Nvidia and AMD support it. Programs like Octane that are very invested in GPU I think could go either way. (adopt openCL, or move to it completely, or abandon OSX completely) Probably depends case by case on the size of the developer for their willingness and the resources they have for change.
^^^^IMHO, really doesn't matter whether the nMP has an Nvidia option or not.
1. Nvidia writes driver and Cuda Updates after each Apple OS Update.
2. Both the MacBook Pro and the iMac have models that use Nvidia Graphics.
3. Apple really doesn't (and hasn't) offered for a while now, reasonably priced Nvidia graphics cards for the current Mac Pro. They still offer the Quadro 4000 (an old card) for $1,200 or the newer Quadro K5000 for $2,500.00. Any one of us own one of those?
So, again, IMHO, both Macintosh and Nvidia will keep drivers and Cuda Updates active for a long time to come.
Lou
1. Nvidia writes driver and Cuda Updates after each Apple OS Update.
2. Both the MacBook Pro and the iMac have models that use Nvidia Graphics.
3. Apple really doesn't (and hasn't) offered for a while now, reasonably priced Nvidia graphics cards for the current Mac Pro. They still offer the Quadro 4000 (an old card) for $1,200 or the newer Quadro K5000 for $2,500.00. Any one of us own one of those?
So, again, IMHO, both Macintosh and Nvidia will keep drivers and Cuda Updates active for a long time to come.
Lou
Would a couple of these work for CUDA in the T/B enclosure?
http://www.amazon.co.uk/EVGA-NVIDIA-Graphics-Card-PCI-E/dp/B00C81T7ZS/ref=sr_1_1?ie=UTF8&qid=1381892463&sr=8-1&keywords=GTX+680+for+mac
http://www.amazon.co.uk/EVGA-NVIDIA-Graphics-Card-PCI-E/dp/B00C81T7ZS/ref=sr_1_1?ie=UTF8&qid=1381892463&sr=8-1&keywords=GTX+680+for+mac
Those drivers are written associated with specific products. For example
http://www.nvidia.com/object/macosx-313.01.02f01-driver.html
on the supported products tab is:
GTX 680 for Mac
GTX 285 for Mac
GT 120
8800 GT
K5000 for Mac, 4000 for Mac
FX 5600, FX 4800 for Mac
Those are all products that went to market for sale. For example, updates in 10.8.3 arrived around the same time as
https://www.macrumors.com/2013/04/0...tx-680-mac-edition-graphics-card-for-mac-pro/
Coincidence?
If there is no card vendor asking them and Apple isn't asking them then there is little to no demand allocate development resources this to cover a broad spectrum of new cards and implementations.
Nvidia has multiple implementation per major architecture rollout. A GK110 is not a GK104
The normal driver distribution path for embedded Mac graphics is through Apple.
So yes there will be a baseline to work with for the corresponding top end iMac GPU (likely an deeply underclocked desktop model ). However, the support is highly limited and variable.
The next iteration of the MBP/iMac will have another design bake off for embedded GPU. If Nvidia manages to loose both that and the next Mac Pro round of its design bake off .... then new cutting edge card drivers for public consumption aren't going to be a high priority.
Does Apple want Nvidia to loose the design bake offs? Probably no more than they want AMD to loose. Multiple vendors vying for design slot wins helps Apple deliver better products over time. But if Nvidia doesn't bring what Apple wants and at the price they want then Nvidia won't get the win.
[ In the two year iteration before 2012-2013 AMD had wins in MBP/iMac space so it is possible to wrap up all of Mac line up if have right designs at the right time at the right price. ]
Goes back to #1. If limited demand has killed off 3rd card vendor interest then where's Nvidia's motivation? What has typically happened is one vendor takes a stab (recently EVGA) and then firmware and drivers go rogue. This results in the numbers sold and margin is so small (have to use discounts to move the cards) the Apple store doesn't want to bother to stock them.
if there was a demand problem of cards priced "too high" before moving them to a external PCIe enclosure isn't going to make them any more cost effective.
what you want to do it's nonsense.
The GPU that the Mac Pro is gonna have would be more than enough for gaming WOW.
The GPU that the Mac Pro is gonna have would be more than enough for gaming WOW.
Where are you getting that pricing from? You can buy that card for $1600. The same price as the PC version.
Yep, the Nvidia page says they are written for specific products, but the Web Drivers have in fact, worked on all Nvidia cards since os 10.8.3. I have a GK110 (flashed GTX780) and it works quite well in ML with both the Apple Drivers or the Nvidia Web Drivers. Nvidia hasn't published web drivers for Mavericks yet, but again the GK110 is performing quite well with the Apple drivers.
As I said, I would be shocked if any of this support changed. In other words, IMHO, "It Aint Gonna Happen" Nvidia support from both Apple and Nvidia will remain with the platform.
Nvidia's publishing of their Web Drivers is in direct conflict with AMD (ATI) which provides absolutely no Apple support to the consumer. The only recent AMD support came from Sapphire for their Mac Edition HD7950, and if you check their web site, the Apple Driver's listed there are hopelessly out of date.
Lou