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

pszilard

macrumors member
Original poster
May 29, 2011
44
4
Sydney, Australia
Wondering if anyone knows whether the new Mac Pro cpu is soldered or is it in a user replaceable socket?

Secondly, are DUAL Xeon chips supported?

I would love to by a 6 core box now, and then add a second 6 core Xeon cpu later on. Am I dreaming? :)
 
They're only using one socket, so there's no reason to design for 2. They are not going to solder in the Xeons. Some of the higher cto versions are quite expensive, and there's nothing gained from it. Everyone asks the same thing because of prior behavior with the other lines.
 
It looks to be replaceable, which would be ideal for BTO options more accessible. Only a teardown would really confirm this. Same for graphics card.
 
So does that mean that I could buy the 4 core and replace it later with a 12 core?

Probably not as it's not the same L3 cache is in general on the motherboard instead of the CPU. The 12-core used by Apple has I think 30 MB while the 4-core has 10 MB. I might be wrong (I'm far from being the more knowledgeable about those thing), but as the L3 cache is different, it's not compatible. (Or maybe could work in a reduced version ?)
 
Probably not as it's not the same L3 cache is in general on the motherboard instead of the CPU.

There's no reason to think that an L3 cache difference will prevent a CPU swap. Assuming you can easily get to the top side of the chip, and there's no evidence to suggest otherwise, then a swap should be quite doable.
 
Probably not as it's not the same L3 cache is in general on the motherboard instead of the CPU. The 12-core used by Apple has I think 30 MB while the 4-core has 10 MB. I might be wrong (I'm far from being the more knowledgeable about those thing),

For the E5 Xeon 2600/1600 class of CPU packages the following are inside.

1. x86 cores including L1 , L2 , L3 caches
2. Memory controllers ( 4)
3. PCIe v2 controller or 40 lanes

OverviewIVB3dies_575px.png

http://www.anandtech.com/show/7285/intel-xeon-e5-2600-v2-12-core-ivy-bridge-ep
For the future Mac Pros that QPI section isn't hooked to anything. It is just kind of along for the ride (and proactively configured permanently 'off' in all of the 1600 series models. )

L3 on the motherboard or discrete in Intel designs is more than several generations ago. Even Intel's new eDRAM Haswell processors have what can effectively be a L4 cache bundled into the package when present. [ Different die but multiple die packages are not a problem. ]

While the socket is the same across all of these it may be up in the air a bit whether Apple is getting either specially binned parts or modified packages that couple more effectively to their custom thermal solution.
 
Last edited:
So the answer is "maybe". :)

Something to keep an eye out for. As for me, I won't be doing this in a hurry, anyway as I won't have any cash left after buying the new MP. LOL. Plus I am not going to toss a perfectly good cpu in the bin.

Never the less, it is interesting to watch out for. Thanks for all the good feedback.
 
For the E5 Xeon 2600/1600 class of CPU packages the following are inside.

1. x86 cores including L1 , L2 , L3 caches
2. Memory controllers ( 4)
3. PCIe v2 controller or 40 lanes

Image
http://www.anandtech.com/show/7285/intel-xeon-e5-2600-v2-12-core-ivy-bridge-ep
For the future Mac Pros that QPI section isn't hooked to anything. It is just kind of along for the ride (and proactively configured permanently 'off' in all of the 1600 series models. )

L3 on the motherboard or discrete in Intel designs is more than several generations ago. Even Intel's new eDRAM Haswell processors have what can effectively be a L4 cache bundled into the package when present. [ Different die but multiple die packages are not a problem. ]

While the socket is the same across all of these it may be up in the air a bit whether Apple is getting either specially binned parts or modified packages that couple more effectively to their custom thermal solution.

Thanks for your correction. I wasn't aware of that.
 
For the E5 Xeon 2600/1600 class of CPU packages the following are inside.

1. x86 cores including L1 , L2 , L3 caches
2. Memory controllers ( 4)
3. PCIe v2 controller or 40 lanes

.

Thanks for your correction. I wasn't aware of that.

There are 40 lanes of PCIe 3.0 per socket.

PCI Express Revision 3.0
PCI Express Configurations ‡ x4, x8, x16
Max # of PCI Express Lanes 40

http://ark.intel.com/products/75273/Intel-Xeon-Processor-E5-2667-v2-25M-Cache-3_30-GHz

The C60x chipset has an additional 8 lanes of PCIe 2.0. http://ark.intel.com/products/63984/Intel-BD82C602-PCH
 
There are 40 lanes of PCIe 3.0 per socket.

Yes. Sorry, I was thinking v3 but it came out 'v2'.

The C60x chipset has an additional 8 lanes of PCIe 2.0

In most PC designs, the Thunderbolt controller is typically attached to the chipset ( IOhub). But in the new Mac Pro so are the Ethernet ports , wifi/bluetooth , USB 3.0 and perhaps audio. In most designs, those 8 are consumed, if not significantly oversubscribed.

Here in the Mac Pro, there are likely three TB controllers (unless Intel springs some super duper, top secret one out of the blue ). There is no way to attach more than one to the C602 highly diluting the bandwidth.
 
It's quite easy

Here in the Mac Pro, there are likely three TB controllers (unless Intel springs some super duper, top secret one out of the blue ). There is no way to attach more than one to the C602 highly diluting the bandwidth.

Just put the GPUs on PCIe 3.0 x8 slots, and you have plenty of extra slots to use for TB.

You know that's what Apple has done....
 
Just put the GPUs on PCIe 3.0 x8 slots, and you have plenty of extra slots to use for TB.

You know that's what Apple has done....

Are you saying that Apple is using 12 full lanes of PCIe to run the TB, and therefore only 28 are leftover for the GPU? Are you also saying that the GPU are only running at 8X?

I just assumed that the GPU were 16x PCIe (32 lanes used) and the remaining 8 lanes were 3.0, somehow divvied up to cover 6GBps over TB.

Probably wont make a difference in performance, the AMD 7970's in that range have very close to zero benefit from more than 8GBps (although I've only seen gaming benchmarks).

That would mean though that there are potentially 12 lanes of PCIe 3.0 just sitting there unused (40 Lanes - 16 Lanes for GPU - 12 lanes for TB = 12). What a waste of LGA2011.
 
Are you saying that Apple is using 12 full lanes of PCIe to run the TB, and therefore only 28 are leftover for the GPU? Are you also saying that the GPU are only running at 8X?

I just assumed that the GPU were 16x PCIe (32 lanes used) and the remaining 8 lanes were 3.0, somehow divvied up to cover 6GBps over TB.

Probably wont make a difference in performance, the AMD 7970's in that range have very close to zero benefit from more than 8GBps (although I've only seen gaming benchmarks).

That would mean though that there are potentially 12 lanes of PCIe 3.0 just sitting there unused (40 Lanes - 16 Lanes for GPU - 12 lanes for TB = 12). What a waste of LGA2011.

I'm saying that Apple *could* be running one or both GPUs on x8 slots to deal with the over-subscription. (And those x8 slots are just as fast as the x16 slots in the CheeseGrater Pro.)

Also, there are 8 lanes of PCIe 2.0 - which would be fine for T-Bolt.
 
Last edited:
Just put the GPUs on PCIe 3.0 x8 slots, and you have plenty of extra slots to use for TB.

You know that's what Apple has done....

No, I don't know. Frankly, if there wasn't a PCIe SSD it wouldn't be required. Anyway, there is absolutely no reason to drop both GPUs down to x8.

...
Also, there are 8 lanes of PCIe 2.0 - which would be fine for T-Bolt.

Pragmatically, there are not. The Mac Pro has two Ethernet sockets. USB 3.0 ports , and Wifi/Bluetooh. All of those eat into the PCIe v2.0 lane budget.

What is included in the system is

2 Ethernet ports ===> two 1x PCIe v2.0
USB 3.0 ===> one 1x PCIe v2.0
Wifi/Bluetooth ===> one 1x PCIe v2.0

That's is four PCIe v2 lanes right there. ( presuming Audio is fully leveraged off the chipset and won't need one. Otherwise pick one of the above to split on a switch. ).

Add a TB controller (x4) and the chipset's lanes are all gone. [ can play games by adding PCIe SSD here but still all gone if it is a x4 v2 device ]


Still have two TB controllers. E5 Xeon has 40. So.

GPU daughtercard ==> x16 PCIe v3.0
GPU daughtercard ==> x16 PCIe v3.0
Thunderbolt Controller ==> two x4 PCIe v2.0

That's 40. It would actually all fit. But the the SSD runs that over.
If it is a x4 PCIe v2 SSD then one of the four x4 PCIe v2.0 has to pair up or need to "borrow" from one of the x16 bundles.

If there is a switch which can trunk down x4 PCIe v2.0 into x2 PCIe v3.0 worth of utilization then don't really need to gut the GPUs. Likewise a switch that doesn't drop all sides down to v2.0 if just one is. So for example a:

x16 PCI v3.0 by ( x16 PCI v3.0 by x4 PCIe v2.0 ) switch. (probably somewhere near x10-12 throughput depending upon how well juggle around the x4 v2.0 )

Or

x4 PCIe v3.0 by ( x4 PCI v3.0 + x4 PCIe. v2.0 ) switch.

Both would have less than x8 impact on the GPU cards. Although in the second case the SSD and one of the thunderbolt networks would have major impact in concurrent use.

Worst case the E5 is configured x16 , x8 , x8 , x4 , x4 And just use a

x8 PCIe by ( x4 PCI + x4 PCI ) switch

to crank out four x4 bundles ( leaving door open to perhaps have dual SSDs in future configurations ). Could move all three TB controllers and SSD here. The vacated chipset lanes could go to connection to power management (typically a PCI bridge ) and other misc internal management duties. ( also audio if necessary). If Apple wanted to maximize USB 3.0 throughput they could use two 2 port/socket USB 3.0 controllers to provision the 4 sockets.


This is far more a technology mismatch. In terms of bandwidth the SSD and TB controllers really only need x2 worth of PCIe v3. If the SSD is a v2 device it is drag on maximum E5 PCIe controller utilization as much as the TB controllers are. A x2 PCIe v3 SSD would put less "borrowing" pressure on the set-up.
 
you dont need a dedicated lane for each ethernet controller. GigEnet is nowhere close to saturating even a first-gen x1 PCIe link.

but it does bring up a good point - until the tech specs are fully explained in detail, there are a lot of potential pitfalls for users on the new mac pro. i think more than ever before in apple history, this is a 1st-gen machine to avoid.
 
you dont need a dedicated lane for each ethernet controller. GigEnet is nowhere close to saturating even a first-gen x1 PCIe link.

but it does bring up a good point - until the tech specs are fully explained in detail, there are a lot of potential pitfalls for users on the new mac pro. i think more than ever before in apple history, this is a 1st-gen machine to avoid.

Exactly, and this is the problem with the TB/USB3 replaces SATA is no problem camp. We don't know where the choke points are.

This whole machine seems set up to come crashing down under heavy load, in general.

What happens when you actually want to use the CPU, both GPUs, the SSD and a TB RAID? Will it keep humming along, or are we going to see I/O saturation and/or throttling due to heat?
 
Exactly, and this is the problem with the TB/USB3 replaces SATA is no problem camp. We don't know where the choke points are.

This whole machine seems set up to come crashing down under heavy load, in general.

What happens when you actually want to use the CPU, both GPUs, the SSD and a TB RAID? Will it keep humming along, or are we going to see I/O saturation and/or throttling due to heat?

You are so silly... Do you think Apple would commit such a childish mistake? Do you have any idea of how many hours this machine has kept being tested, stressfully, for hours and hours of heavy work?

Apple wants to show the world they can continue innovating, so it's not probable that the new mac pro will show faults in its project and cause general public deception.

Just wait and see!
 
You are so silly... Do you think Apple would commit such a childish mistake? Do you have any idea of how many hours this machine has kept being tested, stressfully, for hours and hours of heavy work?

Apple has certainly made mistakes. Appealing to their authority isn't much of an argument against simple skeptisism.

Apple wants to show the world they can continue innovating, so it's not probable that the new mac pro will show faults in its project and cause general public deception.

Just wait and see!

Innovation just for the sake of innovation is really dumb though. There has to be a need that is filled. And frankly. cramming a single CPU, two GPUs and a bunch of TB ports into a small cylander just because you can is hardly innovating. Its just playing with the form factor.

Now if Apple actually makes the machine so artificially imposed bottlenecks aren't reached and makes it easier for developers to utilize the two GPUs in their code, great. I wouldn't get your hopes up though.
 
you dont need a dedicated lane for each ethernet controller. GigEnet is nowhere close to saturating even a first-gen x1 PCIe link.

If you buy a 1x GbE controller you do. The GbE controller you get "for free" in the 602 chipset 87579LM

http://www.intel.com/content/www/us/en/chipsets/server-chipsets/server-chipset-c600.html

only does one port

http://www.intel.com/content/www/us...82579-gigabit-ethernet-connection-family.html

Apple could be by-passing this low power, smaller, cost effective solution for a single dual port controller. They certainly are by-passing all of the SATA and USB 2.0 capabilities of the chipset. But if Apple can use same parts here as on other Macs they tend to go with the volume purchase approach.



but it does bring up a good point - until the tech specs are fully explained in detail, there are a lot of potential pitfalls for users on the new mac pro. i think more than ever before in apple history, this is a 1st-gen machine to avoid.

Apple is going with "tick" CPU processor (v2) that is more process shrink than new micro-architecture . The chipset is exactly the same. So unlikely latent bugs laying around.

Thunderbolt 2 is a bit bleeding edge in terms of traffic management and QoS implementation, but it is also version 2. It is the 3rd generation TB controller.

The W7000-W9000 FirePro cards the Apple versions are dervived from hit the market in June 2012. Again bugs, serious glitches, problems are all likely quite documented at this point.

The USB 3.0 discrete controller is likely 2nd (or 3rd) generation.


The thermal management system is new but picking direct data for the feedback loop likely will work quite well. "Clever" indirect approach vs. straight measure RPMs :

http://www.anandtech.com/show/7501/amd-changes-290-series-fan-algorithms


It isn't like Apple "rushed" this product to market. It is version 1 so there are probably a couple of things that got overlooked. The design seems a little bit ahead of the technology ( too PCIe v2 and Thunderbolt 2 was targeted for 2014 ). But not too far from other version 1 Macs that got a major upgrade.
 
No, I don't know. Frankly, if there wasn't a PCIe SSD it wouldn't be required. Anyway, there is absolutely no reason to drop both GPUs down to x8.

I do wonder about that odd location for the PCIe SSD being right on one of the GPU card's. Maybe we are seeing the other card being run 16x and the one with the SSD being split 8x for the GPU and some other amount of lanes for the SSD, I guess it would need at least 2 lanes for 1.25GB/s, maybe more for future upgrades.
 
[G5]Hydra;18356652 said:
I do wonder about that odd location for the PCIe SSD being right on one of the GPU card's.


The SSD needs to be relatively easily accessible because they wear out. Over an extended length of time and aggressively used, the SSD will likely, relative to the more solid state components, fail. The "remove old, replace with new card" task should be straightforward to do.


It is not odd at all given all the rest of the components in the Mac Pro. Largest inside is the internal triangle of the thermal core and the three major logic/circuit cards.

The card is going to oriented vertically because everything inside is oriented vertically.


On CPU card? buried. One side facing core. Other side facing power supply. There is no way to get to the card to insert a PCIe SSD. On the other side of power supply is the IO controller card and the external physical ports. Again no room for easy access PCIe card socket.

The DIMMs are in the somewhat at the triangle edges.

So the only major open surfaces to put a PCIe SSD socket on is the backs of the GPU cards.



Maybe we are seeing the other card being run 16x and the one with the SSD being split 8x for the GPU and some other amount of lanes for the SSD,

A good chance the connectors on the cards are the same. It is just cheaper to buy at a higher volume and perhaps just not connect all the pins on some boards.


I guess it would need at least 2 lanes for 1.25GB/s, maybe more for future upgrades.

Using x4 v2 lanes now but the base connector/backplane design being able to move to x4 v3 lanes would leaves tons of headroom for future systems, not necessarily upgrades on this specific generation.

Most likely not going to see massive SSD speed increases out of this specific implementation's infrasttructure. 1200MB/s to 1600MB/s perhaps (33%) but nothing like the SATA Flash drive -> PCIe drive move made here.
 
I thought I could clearly see the SSD PCIe card on one of the photos on the MacPro publicity site. That means that a change should be dead easy. Look at the Storage tab.
 
Yes, you just have to have something to replace it with.

I have a 2012 rMBP.

OWC sells replacement SSDs but it seems only in 256 and 512 size. It apparently uses same one as an Air. Oddly, more than a year after intro it seems the only ones to offer a 768 are ones yanked from Apple machines.

So used ones go for $1K.

Lack of competition leads to high prices.
 
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.