Graphics cards and the New MacPro

[MacGamver]

As for the landfill comment... You're absolutely right. However, that article on SSD life expectancy I linked to above gave the Samsung SSD a life span of somewhere between 75 and 200 years depending on usage... So if your great great grand children are still using your 2013 mac Pro around 2100, they can probably still be confident in their data integrity.

Ha ha! Maybe my great-granchildren *might* want to use my MacPro Turbine, but alas, on the 124 and a half year, it has hopelessly failed!

 

[Krevnik]

Which motherboards? I've never heard of passing Display Port over PCIe... The ones I've seen either use on-die GPU (option 1) or a cable (option 2).

Demonstration of the "headless" functionality: http://www.youtube.com/watch?v=O1t7Rc9qFgI

The card in the demo is only connected to the motherboard via PCIe and power. But note that I didn't say "Display Port over PCIe", but "display information". Which is exactly what is being done in that case, using Lucid MVP as the specific technology in order to do it.

The reason to use bridge cables is to avoid clogging the PCIe bus with extra data, especially useful in SLI when the bus is already saturated, or in the case of thunderbolt, to avoid trying to pass a DP signal over PCIe if you want to tap into the card's DP output directly. But it isn't required if you are going to move the frame buffer itself to the integrated GPU which then just flips buffers for you. If you can do this mostly through DMA, you can keep the overhead manageable, and you have to deal with this in some manner via SLI to begin with.

One thing that occurs to mind, is that if you can DMA straight into the integrated graphics' frame buffers, it wouldn't be too bad at all. You'd need to do a little management of the DMA at the driver level, but otherwise your overhead would be pretty low (although at the cost of a little latency).

Considering Apple's GPU firmware is already custom at this point, doing a couple tweaks to reduce overhead of using the integrated GPU as the frame buffer in an SLI situation wouldn't be terribly surprising to me. PCIe 3.0 also provides a bit more buffer to prevent the overhead from becoming an issue (and lessening the need for an SLI bridge cable).

 

[VirtualRain]

Demonstration of the "headless" functionality: http://www.youtube.com/watch?v=O1t7Rc9qFgI

The card in the demo is only connected to the motherboard via PCIe and power. But note that I didn't say "Display Port over PCIe", but "display information". Which is exactly what is being done in that case, using Lucid MVP as the specific technology in order to do it.

The reason to use bridge cables is to avoid clogging the PCIe bus with extra data, especially useful in SLI when the bus is already saturated, or in the case of thunderbolt, to avoid trying to pass a DP signal over PCIe if you want to tap into the card's DP output directly. But it isn't required if you are going to move the frame buffer itself to the integrated GPU which then just flips buffers for you. If you can do this mostly through DMA, you can keep the overhead manageable, and you have to deal with this in some manner via SLI to begin with.

One thing that occurs to mind, is that if you can DMA straight into the integrated graphics' frame buffers, it wouldn't be too bad at all. You'd need to do a little management of the DMA at the driver level, but otherwise your overhead would be pretty low (although at the cost of a little latency).

Considering Apple's GPU firmware is already custom at this point, doing a couple tweaks to reduce overhead of using the integrated GPU as the frame buffer in an SLI situation wouldn't be terribly surprising to me. PCIe 3.0 also provides a bit more buffer to prevent the overhead from becoming an issue (and lessening the need for an SLI bridge cable).

Yeah, this has come up before... but it's a software kludge... AND it requires an IGP on the CPU for the frame buffer remapping. It's simply not possible on a Xeon without an IGP.

 

[Radiating]

Hey all!

I couldn't seem to locate the answer to this question: The new MacPro will have dual PCI-e GPUs, but is this a custom, proprietary slot? What is the likelihood of being able to purchase a third party card that will work? (drivers aside) Is Apple shooting themselves in the foot again by trying to create their own "standards" for things?

The pin-out for the card is identical with no differences whatsoever to the current PCI express 3 standard. The only thing that Apple changed was the CONNECTOR, so that they could squeeze everything into a smaller footprint, a 1 dollar cable will easily let you add any video card you could ever want to the system, driver's aside.

 

[Krevnik]

Yeah, this has come up before... but it's a software kludge... AND it requires an IGP on the CPU for the frame buffer remapping. It's simply not possible on a Xeon without an IGP.

Which is fair to point out. I do wonder how they would handle a proprietary connector if they are doing SLI bridging or DP output over the connector. It's a fair number of pins to bring out 3 DP outputs from a single card.

 

[slughead]

Or option 3) You run the display information over PCIe to the thunderbolt port. This is how the current crop of thunderbolt motherboards do it.

If you do that, you don't really need external connectors for the PCIe GPU, which is what Apple has done. I'd bet that the connection to the mobo is still a standard PCIe pin-out, with a custom board layout to fit the case.

This is only possible if the CPU has an integrated GPU on it. The new Xeons likely wont have this, so it's impossible.

----------

Which is fair to point out. I do wonder how they would handle a proprietary connector if they are doing SLI bridging or DP output over the connector. It's a fair number of pins to bring out 3 DP outputs from a single card.

I'd be very impressed if they had SLI (or, more likely, crossfire) in OS X. As for the amount of pins, they're designing their own connector, they can have as many as they want

 

[slughead]

Thunderbolt requires integration of display port and PCIe. This can only be done by building custom GPU cards OR via some kind of cable kludge. You can guess which one of those Apple would consider a non-starter. Once you go custom GPU, you might as well make it fit your desired form factor.

+1, agreed

According to our own little PCIe slot thread (around here somewhere), the #1 thing people upgrade, by far, was the GPU. Once Apple decided to use proprietary GPU, for most users it may not be worth it to have PCIe slots at all. Apple probably has the hardware profiles of nearly everyone who uses a Mac Pro and knows what configs people have (as well as knows when you are sleeping, are awake, are being bad or good etc).

I suspect Apple looked into the ASUS-type solution and found it both un-Apple-like and probably likely to get rejected by Intel (as ASUS has been).

On a side note, Intel seems hell-bent on making TB a failure by locking everyone who would like the throughput without the motherboard restrictions. We'll see how it goes, but I suspect TB1 and TB2 will be predominantly a phenomenon of the Mac. We'll see if their little add-in card vaporware gets any traction. I could see mobo manufacturers adding GPIO headers and the required bios features but who knows if anyone will buy into it--it'll depend on what hardware they actually want to hook up. There are a ton of motherboards (including my PC) that have Firewire headers on the mobo, in spite of USB3 and other features usurping it. It's probably just a few cents more per motherboard and therefore worth the "THUNDERBOLT READY" bullet point on the spec sheet. When you have display ports, PCIe slots, and internal storage, TB really isn't all that useful....

 

[spaz8]

This is only possible if the CPU has an integrated GPU on it. The new Xeons likely wont have this, so it's impossible.

----------



I'd be very impressed if they had SLI (or, more likely, crossfire) in OS X. As for the amount of pins, they're designing their own connector, they can have as many as they want

Image

Ya I will be happy if we see an Nvidia option in the nMP. I worry it will just be AMD, but perhaps that was just marketing hype that OSX will finally have Firepro card support. OpenCL might be the future, but CUDA is today.

 

[MacGamver]

The pin-out for the card is identical with no differences whatsoever to the current PCI express 3 standard. The only thing that Apple changed was the CONNECTOR, so that they could squeeze everything into a smaller footprint, a 1 dollar cable will easily let you add any video card you could ever want to the system, driver's aside.

Whoa... really? That's interesting, and of course if the thing would physically fit into it, which remains to be seen.

 

[Radiating]

Whoa... really? That's interesting, and of course if the thing would physically fit into it, which remains to be seen.

New video cards won't physically fit at all, Apple uses a special size and lay out for the video card, but that doesn't mean you couldn't use an extension cable and mount them them externally. Someone will have to design an enclosure though because most people won't like exposed video cards hanging out of their computers, preferably an enclosure that attaches to the base unit.

 

[ActionableMango]

The pin-out for the card is identical with no differences whatsoever to the current PCI express 3 standard. The only thing that Apple changed was the CONNECTOR, so that they could squeeze everything into a smaller footprint, a 1 dollar cable will easily let you add any video card you could ever want to the system, driver's aside.

I am genuinely curious, where is this 1-dollar cable that will go from the new Mac Pro's compact connector to standard PCIe3?

 

[Krevnik]

This is only possible if the CPU has an integrated GPU on it. The new Xeons likely wont have this, so it's impossible.

You correct me, and then go on to quote my post where I admit VirtualRain's correction on that matter. Thanks.

I'd be very impressed if they had SLI (or, more likely, crossfire) in OS X. As for the amount of pins, they're designing their own connector, they can have as many as they want

Dual GPUs in this case, if they aren't SLI, aren't terribly interesting outside of OpenCL. You don't exactly need both to drive 3 displays.

The issue is that PCIe already is pretty pin-heavy as it is. Then you are adding 3 DP channels on top of that? Combine that with a more compact connector, and you are pushing the pins on the board pretty close, even ignoring the 12 more lanes you need for the DP data.

I still wouldn't be terribly surprised IMO if they did something custom to pass display data back across the PCIe bus to the TB controller. Although I think a teardown would be needed to demonstrate this either way.

 

[MacGamver]

New video cards won't physically fit at all, Apple uses a special size and lay out for the video card, but that doesn't mean you couldn't use an extension cable and mount them them externally. Someone will have to design an enclosure though because most people won't like exposed video cards hanging out of their computers, preferably an enclosure that attaches to the base unit.

Which connection are we taking about? I'm referring to the PCI-e slot itself... Of course there are already PCI-E chassis' for external usage; and it's been suggested there are/will be ones utilizing TB, if you were going the external route anyway.

 

[MacVidCards]

The pin-out for the card is identical with no differences whatsoever to the current PCI express 3 standard. The only thing that Apple changed was the CONNECTOR, so that they could squeeze everything into a smaller footprint, a 1 dollar cable will easily let you add any video card you could ever want to the system, driver's aside.

There are 3 instances that could explain this post.

1. You are an engineer for Apple working on nMP.

2. You are working for a 3rd party who has a nMP on loan to design peripherals for it.

3. You have pulled this info from the air around you, i.e., made it up.

In any case, I'm sure we are all interested in this $1 cable.

 

[VirtualRain]

Which connection are we taking about? I'm referring to the PCI-e slot itself... Of course there are already PCI-E chassis' for external usage; and it's been suggested there are/will be ones utilizing TB, if you were going the external route anyway.

There's no doubt the new Mac Pro GPUs use a special custom connector that connect the GPU PCB to the bottom PCB for PCIe, power, and display port output although no one will really know for sure until we see it in person (or iFixit gets their hands on one).

I think it's crazy to assume anyone will come up with a solution that adapts a regular PCIe graphics card to this custom connector... and if they do, it certainly won't cost a dollar.

The only potentially viable GPU upgrade solution for the new Mac Pro might possibly be a DIY replacement of the GPU boards in the 2013 model with GPU boards from a future model (eg. 2015 Mac Pro). But even that will depend on no physical mounting, connector, power envelope, or display port output changes between model years. And the cost effectiveness of this will likely be questionable compared to selling it and buying a whole new machine.

Adding additional GPUs via external TB PCIe boxes is not really "upgrading" so much as "expanding" your GPU options. For example, you might want to do this for CUDA support, but that's about the only possible reason this might make sense in my mind.

In general, I think there's three upgrade philosophies you can adopt with the new Mac Pro...

1. Buy a minimal config for minimal investment and flip it every 2 years to get the latest tech. People shouldn't discount this option if the resale values of the new Mac Pro are fairly healthy like other Apple products. You might be able to buy a $3000 config in 2013, sell it for $1500 in 2015, and buy one with the latest GPU and SSD at that time for another net $1500 investment. Getting the latest GPUs (and other advancements) every other year for $1500 isn't terrible. If you're a gamer or otherwise want/need the latest and greatest GPUs in a Mac OS X machine, this is probably your best route (outside of a Hackintosh). You're likely already spending $700-$1000 every two years on a new GPU, SSDs, and other things to keep your system current so spending $1500 every 2 years to have the latest Mac Pro shouldn't be a huge stretch.

2. Buy the best config you can afford in hopes that it's future proof to last you 3-5 years. This is a safe bet if the machine you purchase initially exceeds your computing needs and you're not always feeling a need to have the latest and greatest.

3. Buy something that fits your needs today and hope that DIY upgrades (as mentioned above) become possible with each new refresh of the Mac Pro.

Personally, I'm probably going with #1. How often I flip it will depend completely on how compelling the upgrades are that get added in subsequent model years.

 

[slughead]

In general, I think there's three upgrade philosophies you can adopt with the new Mac Pro...

Totally OT: VirtualRain and I totally disagree on whether or not we "like" the nMP. However, I see his posts again and again as the most realistic, pragmatic, well-founded, and non-BS of just about anyone who thinks the nMP is a positive move for the product line. We totally agree on 90% of the finer points of this machine's advantages / limitations, just disagree on the big picture.

There are 3 instances that could explain this post.
[...]
3. You have pulled this info from the air around you, i.e., made it up.

What was Ockham's Razor again?

Welcome back, by the way.

I am also pretty interested in a $1 adapter for a GPU to the nMP. It'd probably also need an extension (how it would get out of the case, I do not know). Can someone make a photoshop mock-up of Two Nvidia Titan's duck-taped 2 inches off the nMP case connected by a big fat PCIe extension ribbon?

I was just reading about PCIe extension cables on the bitcoin websites. Apparently for low-bandwidth operations (like bitcoin mining) they work pretty well, but there was talk about "catching fire" or "not booting" with the long-ish cables (shorter ones appear fine)... Totally serious, there was fire involved.

 

[VirtualRain]

Totally OT: VirtualRain and I totally disagree on whether or not we "like" the nMP. However, I see his posts again and again as the most realistic, pragmatic, well-founded, and non-BS of just about anyone who thinks the nMP is a positive move for the product line. We totally agree on 90% of the finer points of this machine's advantages / limitations, just disagree on the big picture.

Thanks. It's great to have divergent views and still be able to carry on an intellectual discussion.

 

[Radiating]

There are 3 instances that could explain this post.

1. You are an engineer for Apple working on nMP.

2. You are working for a 3rd party who has a nMP on loan to design peripherals for it.

3. You have pulled this info from the air around you, i.e., made it up.

In any case, I'm sure we are all interested in this $1 cable.

Or, option #4. I'm an engineer and I took a look at a high resolution freely available images of the Mac Pro's demo units from the media conference. Apple encouraged the bare exposed circuit boards of the new machine to be photographed in high resolution and the circuit boards of the video cards are clearly visible with clearly labeled parts and clear circuit paths. The bottom connector on Firepro in the Mac Pro has a standard PCIe 3.0 pin-out and it's LABLED as being PCIe 3.0 literally right there with each connection going to the expected portions of the card.

It would be crazy if nobody made a conversion cable to plug in a standard video card. I've made conversion cables like this in a few hours, it's child's play.

 

[deconstruct60]

Or, option #4. I'm an engineer and I took a look at a high resolution freely available images of the Mac Pro's demo units from the media conference.

I've seen about this close

http://www.engadget.com/gallery/next-generation-apple-mac-pro-uncased-at-wwdc-2013/#!slide=485190


I haven't found even closer photos. Is there a link to another closer collection?

The bottom connector on Firepro in the Mac Pro has a standard PCIe 3.0 pin-out and it's LABLED as being PCIe 3.0 literally right there with each connection going to the expected portions of the card.

But is PCIe 3.0 the only pins on the board's connector? I think it would be slightly more than a $1 cable. Pins for the video output? Power (or this new franken-mac set up will use external power supply for external card. ) d?
Power/Temperature management ( given the fan and heat sink are off the card, it would be surprising if temperature data is being fed into something that is also hooked to those two. )

It would be crazy if nobody made a conversion cable to plug in a standard video card. I've made conversion cables like this in a few hours, it's child's play.

If put the cover back on how does the cable connected? If open to solution space to dismantling the Mac Pro and putting chunks into a new container it is possible. A complete solution for a $1, not particularly. There would be several other pieces/parts that cost money.

 

[slughead]

If put the cover back on how does the cable connected? If open to solution space to dismantling the Mac Pro and putting chunks into a new container it is possible. A complete solution for a $1, not particularly. There would be several other pieces/parts that cost money.

It would be crazy if nobody made a conversion cable to plug in a standard video card. I've made conversion cables like this in a few hours, it's child's play.

Sit back, think it through..

Edit: woops, I thought the titan was 12 inches and the nMP was 8 (they are both around 10")... still going to be ridiculous though.

 

[MacGamver]

Sit back, think it through..

Edit: woops, I thought the titan was 12 inches and the nMP was 8 (they are both around 10")... still going to be ridiculous though.

Heh, I'm not an engineer, I'm an artist... However from the images, the 2 firepro cards look like they are 2 ~8" x ~5" rectangles; nothing like any GPU currently being used now. Since Apple bought an AMD facility, there's no doubt it's custom made for them. I just really think that was a bad move. :/

 

[deconstruct60]

Since Apple bought an AMD facility,

When? First, AMD is a fabless company for several years. The facilities they have are primarily office and research lab space. If talking about cherry picking off employees ( http://appleinsider.com/articles/13...rlando-houses-former-amd-intrinsity-engineers ) that is different and has every indication to have far more to do with iOS than OS X.

there's no doubt it's custom made for them.

The GPU daughtercard is custom but so is the current Apple Mac Pro motherboard. In fact, all of Apple's Mac motherboards are custom. It is simply just what they do. If look at the Mac line up in total, going to an Apple designed solution is simply "normal procedure".

I just really think that was a bad move. :/

If going custom allows Apple to deliver a FirePro card at better prices then they'll will probably swamp a decent number of "that was a good move" folks for "that was a bad move" folks.

 

[MacGamver]

When? First, AMD is a fabless company for several years. The facilities they have are primarily office and research lab space. If talking about cherry picking off employees ( http://appleinsider.com/articles/13...rlando-houses-former-amd-intrinsity-engineers ) that is different and has every indication to have far more to do with iOS than OS X.




The GPU daughtercard is custom but so is the current Apple Mac Pro motherboard. In fact, all of Apple's Mac motherboards are custom. It is simply just what they do. If look at the Mac line up in total, going to an Apple designed solution is simply "normal procedure".




If going custom allows Apple to deliver a FirePro card at better prices then they'll will probably swamp a decent number of "that was a good move" folks for "that was a bad move" folks.

Yes this was the story I was mashing up my details with: https://www.macrumors.com/2013/05/31/more-details-on-apples-new-orlando-gpu-design-center/

And yes of course their stuff is custom designed, but I meant at a more granular level. i.e. while their motherboard may be custom built, the RAM slots themselves are standard. That was more what I meant, rather than creating totally proprietary connections for things. e.g. NuBUS or ADB.

 

[ActionableMango]

If going custom allows Apple to deliver a FirePro card at better prices then they'll will probably swamp a decent number of "that was a good move" folks for "that was a bad move" folks.

It would be extremely out of character for Apple to deliver a video card at better prices than the comparable OEM solution. Even if they could cut costs down, I suspect they would simply keep the savings as margin.

I suppose time will tell.

 

[MacVidCards]

Or, option #4. I'm an engineer and I took a look at a high resolution freely available images of the Mac Pro's demo units from the media conference. Apple encouraged the bare exposed circuit boards of the new machine to be photographed in high resolution and the circuit boards of the video cards are clearly visible with clearly labeled parts and clear circuit paths. The bottom connector on Firepro in the Mac Pro has a standard PCIe 3.0 pin-out and it's LABLED as being PCIe 3.0 literally right there with each connection going to the expected portions of the card.

It would be crazy if nobody made a conversion cable to plug in a standard video card. I've made conversion cables like this in a few hours, it's child's play.

You traced 164 pins and followed the circuits? Through a multilayer PCB ? I'm sorry, that is total BS.

I have repaired damage to MANY GPUs in last 7 years, the traces are barely visible, then vanish from top layer to connect through an intermediate layer and show up 3 inches away from start at a component on backside. And it is in fact the backside that you have pulled this claim from.

Without multilayer schematics and hours tracing circuits there is 100% NO WAY to make any such claim .

I call "shenanigans"

In past Apple changed just a couple pins on AGP connector for use with ADC display and people have had to tape pins 3 & 8 on cards for G4s ever since.

I suggest you recant, or admit it's option 1 or 2.

Or explain how your X-ray vision allows you to see the inner PCB layers from jpegs.