Mac Pro 6,1 Cylinder - Fact Gathering.

[Varmann]

Doubtful. The bulk of the bill of material costs for the current Mac Pro are CPU , RAM , HDD , GPU , power supply , case.

roughly in order of price ( the CPU and GPU flip in non entry configs )

Judged by the build quality of the case I would guess that it moves up a bit on that list. A big case also costs more to ship around. I guess the cases, at a minimum, are shipped around three times before reaching the buyers.

Dual Xeon motherboards with daughtercards are also a pricey concept, that Apple has get rid of.

 

[rGiskard]

Some more info to add - one based on pictures, the other based on previously available Intel documentation:

- A place for a second SSD is shown in the pictures - the actual connector is not there, but you can see the silk screening on the board for it. This points towards two SSDs being a BTO option, but it's nothing confirmed.

Maybe, but it all indications are that the Mac Pro shown to the public was a maxed out model. If Apple planned on offering a BTO for two SSD blades, why only show one to the public?

The more likely explanation for that silk screening on the board is that the same boards are used for both cards to save costs. There are a few other differences between the boards in electrical components as well, suggesting these are not discreet and interchangable video cards.

Perhaps this thread should retitled to "Speculation not directly contradicted by statements from Apple".

I mean, there might be a BTO version with one video card that filles the third side of the triangle with half a dozen PCIe sockets for SSDs. Apple could possibly maybe cram all that in the tube. It's also possible that if one pops off the top cover in the center of the fan, there is a moist, warm hole featuring 9.9 inches of decadent holiness into which lucky Mac Pro owners can sink their manhood. Not likely, but Apple could definitely cram such a thing into the design.

 

[deconstruct60]

There's an asymmetry here I don't get. TB requires a video signal, which is why I've been expecting on board video all along. Six ports, which seems likely to be 3 TB controllers with two ports each. But two graphics cards. If they were separated, how do you wire two separate cards to six ports connected through three controllers?

Easy. You don't. The current Mac Pro comes with video cards with 3 connectors on them. You either

a. wire up 3 ports each from the GPU to some TB controller ( plus one from GPU to the HDMI port.)


b. wire up 2 on one TB controller and split the 'middle' controller with one from either side (plus one from a GPU to the HDMI or both if want to just add one video switch. ) )

c. wire up 4 ports from one GPU to two TB controllers. And 2 ports from one GPU and a another port from that GPU to the HMDI port.


What you are trying to do is a crossbar switch were video is routed anywhere. That is way too much complexity for the extremely marginal gain. Like with current GPU card edges if you want a particular GPU you pick the connector associated with the GPU.





If the cards were separated, then if you had just one monitor plugged in then the other GPU would be for naught, except for computing tasks. Oh they have an HDMI too, which makes it more interesting.

Doesn't seem to make sense unless they combined them into one GPU pool which could be allocated to the TB ports as necessary.

Even a 6 way crossbar switch doesn't make much sense at all for the marginal gain. If can hook up 6 with straight shots to controllers you can certainly hook up 3-4.

----------

Judged by the build quality of the case I would guess that it moves up a bit on that list. A big case also costs more to ship around. I guess the cases, at a minimum, are shipped around three times before reaching the buyers.

Yeah , for base entry unit it is higher proportion. Although when get to higher dual package RAM configs it is getting closer to dropping down, but that is not where trying to limbo under the $2,500 limit.

Even a case that is $400 isn't going move much when delete the cost of new case. And can BOM cost always go back in with a non-completely-stripped down SSD, if it goes "too low".

Dual Xeon motherboards with daughtercards are also a pricey concept, that Apple has get rid of.

Well that actually gets rid of RAM too. And puts the customers on higher density RAM to go higher. If they sold enough units the daughtercard should have paid for itself. The charge back for that was all Mac Pros. The should be selling enough to pay for custom design of the motherboard.

the cost of the additional materials. Phew. A connector and socket? Of of $2,000+ system that isn't going to be a game changer.

 

[deconstruct60]

There are a few other differences between the boards in electrical components as well, suggesting these are not discreet and interchangable video cards.

The most certainly are discrete. They are separate pieces of logic board. Extremely unlikely that is a 'U' shaped single board with the bottom of the 'U'. Even the Apple's thermal solution animation on the site has it three separate units. All three pieces being daughterboards that plug in is going back to the original NeXT Cube where the primary mainboard plugged into Nubus. How they are connected isn't shown but triangle shape suggests there are three pieces here probably of one logical board. ( the GPUs don't interface with the outside directly at all. )

I mean, there might be a BTO version with one video card that filles the third side of the triangle with half a dozen PCIe sockets for SSDs.

extremely not likely. Frankly the system as configured is oversubscribed on PCI-e lanes. There are "half dozen" extra PCI-e lines to assign to another half dozen sockets.

There is a decent chance the PCI-e SSD is mounted on the GPU daughtcard so that use a switch to share x16 lanes between the GPU section of the card and the SSD.

If out of lanes and want to add a second SSD then taking diluting the other x16 going to the 2nd GPU would be a logical choice.

I don't think that would be the nominal, standard configuration though. If the 2nd cards standard duty is as a compute engine probably don't want to dilute its bandwidth at all. Whereas graphics computation piplelines often have compression trick and optimizations to get around constrained PCI-e. triming x4 off of x16 isn't going to slow it down much. Especially at PCI-e v3.0 ( x12 is still effectively x24 worth of v2.0 bandwidth which is higher than any v2.0 GPU card was giving. )

 

[deconstruct60]

Assuming the gpu's are twin AMD FirePro™ W9000's then the power supply will surely be pricier, but nobody will notice, because the 9000 comes at 3400$ apiece!

It is probably still in the same ballpark.

No Dual Xeon Processor ( 2 x95W --> 190 .. only have to do 130W now so got back 50W )
Tossed two PCI-e slots. 75W each mandates --> got back another 150W

No allocation for SATA drives @5W each and 6 bays that's 30W

So total 230W savings which is pretty close to 250-270W


I wouldn't be surprised though if the the two W9000 equiv configuration had a requirement for a E5 2600 and its closer to 95W TDP. Drop another 20W and definately in 250-270W range.

If they can get the entry price down a bit and unit sales up some then this custom power supply should get cheaper. but yeah that is mainly on unit sales reduction and not a massive power reduction.

Frankly, if it is a flop unit sales wise then all the unique parts are going to get more expensive.

 

[Cubemmal]

Easy. You don't. The current Mac Pro comes with video cards with 3 connectors on them. You either ...

Hard to follow this. Can you explain, in plain terms, how you think the video is wired up and allocated?

 

[deconstruct60]

Hard to follow this. Can you explain, in plain terms, how you think the video is wired up and allocated?

The one I think most probable because TB controllers are sensitive to placement so will be juggling distance between GPU and TB controller (option C in original list but other variants should be obvious derivatives of the below) :

GPU1 output 1 --> TB controller 1 input 1
GPU1 output 2 --> TB controller 1 input 2
GPU1 output3 ---> HDMI

GPU2 output 1 ---> TB controller 2 input 1
GPU2 output 2 ---> TB controller 2 input 2
GPU2 output 3 ---> TB controller 3 input 1
GPU2 output 4 ---> TB controller 3 input 2

The groups are split above just as the edge connectors of the GPUs would be segmented on cards that ran output to edge connectors. The user manual could simply state to start using the top two TB ports for video TB and/or DisplayPort chains. HMDI is obviously just doing video output ( or digital audio in wierd corner cases.) If want to engage both GPU cards to split loads for two screens then start using TB ports vertically from the top.




The current TB controllers have a maximum of two DisplayPort v1.1 inputs. I'm not sure if the DisplayPort v1.2 is going to keep that the same or that the pass-through section will need more inputs to get multiples higher than two out on purely DP v1.2 traffic or that DP 1.2 instances just put more on wires already there (but there has to be a limit to that.)


Frankly if GPU2 is primarily used for GPGPU most of the time then those outputs aren't used. If folks only have 1-2 monitors that's is just fine. The DP inputs are there so pass TB compliance testing that controllers have do video out.

In the current macs with two GPUs can have something like

GPU1 output 1 --------|
GPU2 output 1 ----> DP switch --> TB input 1
GPU2 output 2 ----> DP switch ---> TB input 2
GPU1 output 2 --------|

where the switches are synchronized to be all on GPU1 or GPU2. that approach or a more complicated routing approach increases complexity for little gain. In a mobile case may want to switch because the GPUs are asymmetrc ( one much bigger power saver than the other. If far more closely matched, there isn't alot of upside). With more TB controllers the complexity goes up.



Making GPU2 push all of its frame buffer into GPU1 doesn't really make it any more easier. In fact it probably slightly increases trace lengths since can skew the TB controllers on the board slight to be closer to the one they hook to and pack all three relatively close to the physical ports. )

GPU1 output 1 --> TB 1 input 1
GPU1 output 2 --> TB 1 input 2
GPU1 output 3 --> TB 2 input 1
GPU1 output 4 --> TB 2 input 2
GPU1 output 5 --> TB 3 input 1
GPU1 output 6 --> TB 3 input 2
GPU1 output 7 --> HDMI

GPU1 is so loaded down now it has to have farmed out work to GPU2 and still be at high performance.

 

[Tesselator]

The one I think most probable.

And each can be daisy chained for two displays each (say 1080 displays) for a total of 12 displays on the TB2 lines - plus one HDMI? Or??

 

[tamvly]

And each can be daisy chained for two displays each (say 1080 displays) for a total of 12 displays on the TB2 lines - plus one HDMI? Or??

This is misleading, IMHO. I have two monitors on my 3,1 MP, both of which are Dual Link DVI. So daisy chaining doesn't seem possible. So the number of legacy monitors will be limited and dependent on adaptors.

I use the Apple Dual Link - TB adaptor on my MacBook Pro to connect to one of the monitors (NEC 271W which support multiple inputs). Many folks, including me, think this device is not one of Apple's best efforts. So adaptors are part of the equation and there is a dearth of alternatives.

 

[Tesselator]

This is misleading, IMHO. I have two monitors on my 3,1 MP, both of which are Dual Link DVI. So daisy chaining doesn't seem possible. So the number of legacy monitors will be limited and dependent on adaptors.

I use the Apple Dual Link - TB adaptor on my MacBook Pro to connect to one of the monitors (NEC 271W which support multiple inputs). Many folks, including me, think this device is not one of Apple's best efforts. So adaptors are part of the equation and there is a dearth of alternatives.

I'm not really interested in the limitations imposed by user selected goods. Rather, I wish to know the maximum possible given whichever devices allow that maximum. So I guess TB displays of some typical/common resolution?

 

[VirtualRain]

I think we now have added facts related to the new Ivy Bridge Xeon CPUs...
https://forums.macrumors.com/threads/1597637/

 

[tamvly]

I'm not really interested in the limitations imposed by user selected goods. Rather, I wish to know the maximum possible given whichever devices allow that maximum. So I guess TB displays of some typical/common resolution?

Really? OK, that makes this thread considerably less interesting to me (and perhaps other folks). I'd venture that most of us aren't going to start with a blank slate.

 

[Tesselator]

Really? OK, that makes this thread considerably less interesting to me (and perhaps other folks). I'd venture that most of us aren't going to start with a blank slate.

Yup, this isn't meant to be one of those "fun and interesting" threads where everyone piles on their hopes and fears, their aspirations and expectations. It's an effort to try and get the machine specifications right so folks can have a more solid idea of what's coming and thus how to plan their system should they decide to upgrade. Experienced folk with an analytical mind know most of this just from watching the conference video but by reading this forum I gather they are not the reader majority. And thus this thread. It also should via the revealed specifications, inform on how to make best use of currently owned peripherals so your "blank slate" comment kinda misses the mark there.

 

[goMac]

It is designed for developers of software for Apples other systems which will use GPGPU, and for power users of commercial software written for Apple systems. It is not designed as a general purpose workstation.

This isn't necessarily true...

I think the part that is true is that Apple wants to promote GPGPU. This isn't unreasonable. A lot of pro software is OpenCL (or will be OpenCL shortly), and you can get more bang out of your buck from a GPU. Software that is CPU based only is becoming more rare. And even though, software like Photoshop that is still mostly CPU based can't scale beyond a few cores anyway.

But for developers? This system is way overkill. A 7970 would be something a developer would look at, not a FireGL. A FireGL's ECC memory is of little use to a developer.

In fact, there isn't a lot of advantages this machine would bring to the table for a developer over an existing Macbook Pro for GPGPU. Apple has been supporting multiple OpenCL targets since the dual GPU 2009 Macbook Pro. So dual GPUs on a Mac for OpenCL is nothing new.

 

[scottrichardson]

Can you please include a changelog at the bottom of the original post so that we can easily see what new information has been added and when? Would be great!

Kind regards,

Scottie

 

[Jhowland]

It appears to me that the PSU is mounted to the removable port panel.

 

[Gonk42]

But for developers? This system is way overkill. A 7970 would be something a developer would look at, not a FireGL. A FireGL's ECC memory is of little use to a developer.

In fact, there isn't a lot of advantages this machine would bring to the table for a developer over an existing Macbook Pro for GPGPU. Apple has been supporting multiple OpenCL targets since the dual GPU 2009 Macbook Pro. So dual GPUs on a Mac for OpenCL is nothing new.

It depends what you are developing, I guess. If you are writing systems designed for massive data sets then you need to be able to try them out. I would have thought that if your customers need a lot of GPU memory then you need a lot as well if only to test that things scale ok.

Though I agree, much of the code development could be done on much less powerful hardware.

My earlier comments were really based on GPGPU programming carrying an overhead in terms of added design and programming complexity, even if you have existing code libraries.

For a commercial developer the overhead might be acceptable given a large enough market.

For those developing code for their own use or for smaller markets, or those developing code that doesn't fit the GPGPU model well, they are stuck with hardware that is GPGPU biased with restriction on the number of CPUs (1) and the number of RAM DIMMs (4) whilst an expensive second GPU is under-utilised.

 

[deconstruct60]

And each can be daisy chained for two displays each

There is nothing in Thunderbolt that says that each chain has to have two. In fact, the daisy chains off two host ports are not entirely independent. That is the flawed assumption that more than a few bring to Thunderbolt from I'm not sure where. Not particularly necessarily true for USB or FW or pretty much anything else either that is commonly used to plug things in.

If the TB controller has two DisplayPort inputs then the max is 2.

In a pure DisplayPort v1.2 bypass mode, that may be different depending upon how the GPU is feeding the physical DP v1.2 connection.

When transported by TB to a remote device DP v1.2 signal is probably capped to a smaller fixed number of displays, like TB v1.0's two displays.
It is going to be difficult to use the same set of wires that DP v1.2 maxes out to move video and something else of pumping the same set of maxed out video data down them in a slightly different format.

 

[Tesselator]

Can you please include a changelog at the bottom of the original post so that we can easily see what new information has been added and when? Would be great!

Kind regards,

Scottie

OK, I'll try to do so in the future. Good idea, thanks!

It appears to me that the PSU is mounted to the removable port panel.

It looks to me to be in the bottom base tray:

There is nothing in Thunderbolt that says that each chain has to have two. In fact, the daisy chains off two host ports are not entirely independent. That is the flawed assumption that more than a few bring to Thunderbolt from I'm not sure where. Not particularly necessarily true for USB or FW or pretty much anything else either that is commonly used to plug things in.

If the TB controller has two DisplayPort inputs then the max is 2.

In a pure DisplayPort v1.2 bypass mode, that may be different depending upon how the GPU is feeding the physical DP v1.2 connection.

When transported by TB to a remote device DP v1.2 signal is probably capped to a smaller fixed number of displays, like TB v1.0's two displays.
It is going to be difficult to use the same set of wires that DP v1.2 maxes out to move video and something else of pumping the same set of maxed out video data down them in a slightly different format.

OK, so what should I put for that sentence to make it more correct?

 

[Cubemmal]

I'm a developer and I want to get a new Pro to explore using GPU's/OpenCL more in my software. Clearly the future is in GPU parallalism and we have been slow to take advantage of it. The best AI techniques presently are ANN's and we have recently figured out how to GPU parallelize them for example. I'd like to work on that.

But if you're just writing some crAPPastic software a mini will work well enough.

 

[goMac]

For those developing code for their own use or for smaller markets, or those developing code that doesn't fit the GPGPU model well, they are stuck with hardware that is GPGPU biased with restriction on the number of CPUs (1) and the number of RAM DIMMs (4) whilst an expensive second GPU is under-utilised.

It's true that the Mac Pro shipping single CPU is a drawback. But the market for 24 cores is a pretty slim market. I can only think of a few industries that can use 24 cores, and even then, you could code for 24 core machines from a 12 core machine (not that it maters if Apple isn't making a 24 core machine anyway.)

I could certainly understand why some markets would be mad that Apple is dropping dual CPUs, but I don't think development will be one of them.

I also think Apple will probably have a low end GPU option for people like developers who aren't doing much GPGPU.

 

[deconstruct60]

It looks to me to be in the bottom base tray:

The circuits at the bottom are probably the minimized base backplane that the three major logic cards plug into.

The "cheese grader" thing here

in the thermal solution image that is stuffed between the CPU+C602 (core chipset ) board and the port connetors backplane board is probably the power supply unit.

It is mounted vertically so that it too can leak heat up through the cylinder at very low fan speed and have air pulled through it at nominal and higher fan speeds.

To be mounted horizontally would be extremely dubious when the airflow though the box goes perpendicular to that. Even more so when drawing air from bottom to top of the box to place at heat source at the bottom. You do NOT want to put something in the bottom to "pre heat" the air before going over the heat exchanger unless you snored you way through your thermodynamics class.

Most of heat from the power supply will not go through the CPU board to the heat sink. It should be shot straight up to avoid that whole subassembly all together. The picture's lower set of arrows are a bit dubious in that respect.


What is on the bottom of the case is probably largely passive electronically with extremely low TDP requirements.

OK, so what should I put for that sentence to make it more correct?

The number of monitors supported per Thunderbolt controller is dependent upon how the DisplayPort v1.2 connectivity is set up. Minimally there are 2 monitors per TB controller possible (no worse than current TB v1 solutions).

From these pictures and released info there isn't a good way to see how the DP v1.2 capable controller is set up, but it is extremely unlikely TB v2 backslides from TB v1 two inputs on the higher end (non dongle focused ) host controllers.

 

[zaphoyd]

Does anyone have any information or educated guesses about how DisplayPort 1.2 MST will work with Thunderbolt 2? Does a chain of 2-3 small MST displays work like a single DisplayPort source? I.e. if a Thunderbolt controller allows 2 DisplayPort inputs would that mean 2 MST streams of 1-4 displays each?

Also, someone more familiar with how PCIe works:
- Thunderbolt advertises 20Gbit. That is 2.5GByte/sec. 4x PCIe 2.0 is ~2GByte/sec. Does this imply that Thunderbolt cannot hit its maximum speeds with PCIe data alone (i.e. requires some Displayport data)
- If a Thunderbolt controller is a PCIe 2.0 4x device, can you put one of them on two PCIe 3.0 lanes? Or do you have to put it on 4 lanes that then get clocked down to 2.0 speeds?
- How much overhead is there with PCIe switching? Could a pair of Thunderbolt controllers (4x PCIe 2.0) share a 4x PCIe 3.0 lane without interfering with each other too much?
- Is there any benefit to giving a single FirePro W9000 16x PCIe 3.0 lanes over 8x PCI 3.0 lanes (~equivalent bandwidth to 16x 2.0)
- How much PCIe bandwidth does a typical PCIe SSD use? It seems if Apple is claiming 1.25GB/sec from the SSD that it would need at least 2x PCIe 3.0 lanes? Does that sound right?

 

[Michaelgtrusa]

These look good.

 

[Tesselator]

I added a video and made some changes yesterday.