an 8-core 3.5GHz processor ..., and supports up to 1TB of 2666MHz memory. 12-core chip is clocked at 3.3GHz, ... it also supports up to 1TB of memory the 16-core 3.2GHz ...it also includes support for 1TB of 2933MHz memory. the 24-core 2.7GHz ...the 24-core model can handle up to 2TB of 2933MHz memory, though Apple rates the Mac Pro to handle 1.5TB at this time. Lastly, the high-end 28-core processor is clocked at 2.5GHz, ...with the same memory capacity as the 24-core model.
So if some models are limited to 1TB and some 1.5TB, does that mean the max memory per slot is lower for the 1TB versions, or is the number of usable memory slots reduced for the 1TB versions?
We've been told that the motherboards are the same, and it is a total memory addressing situation, not a per-chip limit.
This is soooo reminiscent of the old cheese grater models in some ways - I'm getting verklempt. Except, this one is better, with that all around easy access and double sided MOB. The old grater Mac Pros were wonderful, but taking them apart was a real pain if you were winging it (i.e. not trained in Mac Pro repair). Getting at fans, dvd drives etc. was very difficult. This new model has outstanding access to almost everything it seems. Creative and useful design work. Too bad I'm retired and no longer have any contact with Mac Pros
And this is the first thing I thought of when I saw the new design. Embrace the cheese grater! Honestly, I really like the new look.
I'm really looking forward to the iFixit article where they buy a $50,000 computer, and then forensically tear it apart. Including that $1000 stand. I want to see o-ring and hinge analysis.
an 8-core 3.5GHz processor ..., and supports up to 1TB of 2666MHz memory. 12-core chip is clocked at 3.3GHz, ... it also supports up to 1TB of memory the 16-core 3.2GHz ...it also includes support for 1TB of 2933MHz memory. the 24-core 2.7GHz ...the 24-core model can handle up to 2TB of 2933MHz memory, though Apple rates the Mac Pro to handle 1.5TB at this time. Lastly, the high-end 28-core processor is clocked at 2.5GHz, ...with the same memory capacity as the 24-core model.
So if some models are limited to 1TB and some 1.5TB, does that mean the max memory per slot is lower for the 1TB versions, or is the number of usable memory slots reduced for the 1TB versions?
We've been told that the motherboards are the same, and it is a total memory addressing situation, not a per-chip limit.
I'm not sure what you mean - it's absolutely a per-chip limit. Intel prevents using more than X amount of RAM with its chips. X depends on the specific chip model. In the last two generations of Xeons, most chips have a certain limit (768GB in the first gen, for example), and "M" models that were otherwise identical to non-M models could address more memory (1.5TB in the first gen), at a much higher price - thousands of dollars more, typically.
These Macs appear to be using unannounced second-gen chips. We'll know for sure when they ship (or explain better).
As I mentioned elsewhere, chances are they're quoting 1.5TB max because they haven't qualified any 256GB DIMMs yet. It's possible that they might later. Each such DIMM would cost more than the entire base Mac, especially if you bought them from Apple!
The Xeons are likely Xeon W-3xxx series CPUs, the Cascades Lake version. Think Pentium D. They are essentially dual-dies. I’m not sure if it is two dies on the wafer right next to each other, or most likely, 2 separate dies put together in the same package. Has got to be the latter for cost and yield reasons alone. With this, Intel has an upgrade path to 32, 48, 56 cores by just using higher core count parts.
I won't swear to it, but I'm almost certain your statement about dual-die is incorrect. You are thinking of the 92xx series, which have 48 and 56 cores, with two dies in one socket. They're fairly crazy chips with very specific and limited use cases. The Mac's Xeon W chips are not like that. They are standard Intel HCC-type Xeons.
These Xeons have 64 lanes of PCIe. It has
8 PCIe slots, plus 2 custom slots with PCIe for the MPX modules. So,
it’ll have to be something like 16-16-8-8-4-4-4-4 for the 8 slots
(something like this), and, they have to use some PCI lanes off the PCH
for 2 TB3 controllers, and hopefully it has native support for 2 10G
Ethernet ports. And I’m never clear on how the display signaling is
split to the TB3 controllers and if those or independent lanes of PCI e or not.
Yes, they have 64 lanes of PCI, but everything else you said is almost certainly wrong (again, no way to know for sure until somebody gets a closer look at the MB). Doing what you describe would be extremely unwieldy and produce terrible bottlenecking. As with all the other server MBs with tons of IO, they likely have one or more PCIe switch chips (often called "PLX"). Running all their TB3 out of the PCH would be ridiculous. In fact, it would be impossible - there is not enough bandwidth in the PCH to begin to support what 12 TB3 links could demand, by an order of magnitude.
With the PCIe switches, all bets are off on slot configuration. I thought they said what they support when they introed it - maybe 4x16, 3x8, 1x4? Mike, do you remember?
an 8-core 3.5GHz processor ..., and supports up to 1TB of 2666MHz memory. 12-core chip is clocked at 3.3GHz, ... it also supports up to 1TB of memory the 16-core 3.2GHz ...it also includes support for 1TB of 2933MHz memory. the 24-core 2.7GHz ...the 24-core model can handle up to 2TB of 2933MHz memory, though Apple rates the Mac Pro to handle 1.5TB at this time. Lastly, the high-end 28-core processor is clocked at 2.5GHz, ...with the same memory capacity as the 24-core model.
So if some models are limited to 1TB and some 1.5TB, does that mean the max memory per slot is lower for the 1TB versions, or is the number of usable memory slots reduced for the 1TB versions?
We've been told that the motherboards are the same, and it is a total memory addressing situation, not a per-chip limit.
I'm not sure what you mean - it's absolutely a per-chip limit. Intel prevents using more than X amount of RAM with its chips. X depends on the specific chip model. In the last two generations of Xeons, most chips have a certain limit (768GB in the first gen, for example), and "M" models that were otherwise identical to non-M models could address more memory (1.5TB in the first gen), at a much higher price - thousands of dollars more, typically.
These Macs appear to be using unannounced second-gen chips. We'll know for sure when they ship (or explain better).
As I mentioned elsewhere, chances are they're quoting 1.5TB max because they haven't qualified any 256GB DIMMs yet. It's possible that they might later. Each such DIMM would cost more than the entire base Mac, especially if you bought them from Apple!
We're saying the same thing. It is not a per-chip limit based on the memory chip capacity with X supported on one an 2X not. It is a total address space one, limited by the CPU & architecture.
The one thing I’m disappointed about is that it’s PCI-E 3. Not 4. As PCI-E 4 motherboards are coming out, and both AMD and Intel announced support for some of their latest high end chips, I would have thought that this would be that. I really want to buy this, this year. I don’t want to buy another one next year, or the one after that. This will have to last me four years. So even if it’s possible to upgrade the CPU, and the GPU(s), I’d hate to be thinking that a double speed bus will be out a year after I bought this, with all the major performance, security and feature enhancements that 4 will bring. Then, PCI-E 5 is expected for 2021. So round we go again.
Does anyone know if it is PCI 4 or 5? There were some rumors about 4 being a gap solution and that most vendors would probably go right to 5 ...
I read somewhere the new MP used 3, but I think/hope that was a mistake.
Later this year, both AMD and Intel will have some CPUs ready for 4, but not right now. From what I remember, Intel will have some Xeons. But it’s possible that it’s too late.
4 is just becoming available in some mobocracy. But even there it’s not all 4, but a hybrid of 3 and 4. 5 isn’t expected until 2021. The problem we’re seeing is that 3 has been around too long.
an 8-core 3.5GHz processor ..., and supports up to 1TB of 2666MHz memory. 12-core chip is clocked at 3.3GHz, ... it also supports up to 1TB of memory the 16-core 3.2GHz ...it also includes support for 1TB of 2933MHz memory. the 24-core 2.7GHz ...the 24-core model can handle up to 2TB of 2933MHz memory, though Apple rates the Mac Pro to handle 1.5TB at this time. Lastly, the high-end 28-core processor is clocked at 2.5GHz, ...with the same memory capacity as the 24-core model.
So if some models are limited to 1TB and some 1.5TB, does that mean the max memory per slot is lower for the 1TB versions, or is the number of usable memory slots reduced for the 1TB versions?
We've been told that the motherboards are the same, and it is a total memory addressing situation, not a per-chip limit.
I'm not sure what you mean - it's absolutely a per-chip limit. Intel prevents using more than X amount of RAM with its chips. X depends on the specific chip model. In the last two generations of Xeons, most chips have a certain limit (768GB in the first gen, for example), and "M" models that were otherwise identical to non-M models could address more memory (1.5TB in the first gen), at a much higher price - thousands of dollars more, typically.
These Macs appear to be using unannounced second-gen chips. We'll know for sure when they ship (or explain better).
As I mentioned elsewhere, chances are they're quoting 1.5TB max because they haven't qualified any 256GB DIMMs yet. It's possible that they might later. Each such DIMM would cost more than the entire base Mac, especially if you bought them from Apple!
I believe that’s correct. I was to,d that 256 wasn’t yet certified. Apple is very conservative. I don’t remember a time when their stated max was correct, when it was less that the addressed amount the chip could see, except for one time, when an engineering error on their part didn’t have all the traces on the mobo. Oops!
I won't swear to it, but I'm almost certain your statement about dual-die is incorrect. You are thinking of the 92xx series, which have 48 and 56 cores, with two dies in one socket. They're fairly crazy chips with very specific and limited use cases. The Mac's Xeon W chips are not like that. They are standard Intel HCC-type Xeons. ...
Yes, they have 64 lanes of PCI, but everything else you said is almost certainly wrong (again, no way to know for sure until somebody gets a closer look at the MB). Doing what you describe would be extremely unwieldy and produce terrible bottlenecking. As with all the other server MBs with tons of IO, they likely have one or more PCIe switch chips (often called "PLX"). Running all their TB3 out of the PCH would be ridiculous. In fact, it would be impossible - there is not enough bandwidth in the PCH to begin to support what 12 TB3 links could demand, by an order of magnitude.
With the PCIe switches, all bets are off on slot configuration. I thought they said what they support when they introed it - maybe 4x16, 3x8, 1x4? Mike, do you remember?
I wouldn’t be surprised if I was wrong with the MCM packaging. It takes a rather large amount of time to keep track of Intel hardware, something I really don’t have time for. Intel also doesn’t make it easy with ARK being with obtuse. Really didn’t think they would put 64 PCIe lanes off the CPU just yet (on a SkyLake derivative), and dual die would be a cheap way for them to do it.
There should be about 84 PCIe lanes total with 64 off the CPU and 20 off the PCH. I don’t know what bandwidth or PCIe lane equivalent of the PCH connection to the CPU. If it is 8x, I can see those 20 used for the single width slots. If it is 4x, maybe the slot used for IO, 10G Ethernet ports, the TB controller for the top ports and the T2, and they’ll need a switch if they are advertising 16x support for the 7 full length slots.
If there is a switch, we will have to wait and see when someone decides to put say 6 Vega cards in it. There’s going to be a penalty.
The one thing I’m disappointed about is that it’s PCI-E 3. Not 4. As PCI-E 4 motherboards are coming out, and both AMD and Intel announced support for some of their latest high end chips, I would have thought that this would be that. I really want to buy this, this year. I don’t want to buy another one next year, or the one after that. This will have to last me four years. So even if it’s possible to upgrade the CPU, and the GPU(s), I’d hate to be thinking that a double speed bus will be out a year after I bought this, with all the major performance, security and feature enhancements that 4 will bring. Then, PCI-E 5 is expected for 2021. So round we go again.
This was never going to happen. Next-gen AMD CPUs (Zen 2/Ryzen 3xxx) and their chipset (570) have announced support for PCIe4, but they are not shipping until 7/7. Intel has not announced *any* PCIe4 platforms/chips for this year. The earliest CPU with PCIe4 support will be "Ice Lake" Xeons sometime next year... assuming they can ship it on time. It's on the "10nm+" process, and I have to say, the first 2019 10nm chips are disappointing. We don't really know if they have finally beaten their yield problems either - though if they haven't, their top executives will wind up in court (and possibly jail), given the things they've been saying to investors.
The only chance we had was if Apple announced that they'd be shipping Ryzens (or next-gen TRs/EPYCs). I would have loved to see that - they're dramatically superior to the current Intels in pretty much every way - but I understand that they're especially risk-adverse at the moment. Maybe next year (would love that, but doubt it).
Later this year, both AMD and Intel will have some CPUs ready for 4, but not right now. From what I remember, Intel will have some Xeons. But it’s possible that it’s too late.
4 is just becoming available in some mobocracy. But even there it’s not all 4, but a hybrid of 3 and 4. 5 isn’t expected until 2021. The problem we’re seeing is that 3 has been around too long.
Intel will not have any this year. AMD's are shipping 7/7. I'm not sure what you mean by "mobocracy" but all the 570 chipset mobos for AMD that I've seen (a couple dozen at least) are pure PCIe4.
I won't swear to it, but I'm almost certain your statement about dual-die is incorrect. You are thinking of the 92xx series, which have 48 and 56 cores, with two dies in one socket. They're fairly crazy chips with very specific and limited use cases. The Mac's Xeon W chips are not like that. They are standard Intel HCC-type Xeons. ...
Yes, they have 64 lanes of PCI, but everything else you said is almost certainly wrong (again, no way to know for sure until somebody gets a closer look at the MB). Doing what you describe would be extremely unwieldy and produce terrible bottlenecking. As with all the other server MBs with tons of IO, they likely have one or more PCIe switch chips (often called "PLX"). Running all their TB3 out of the PCH would be ridiculous. In fact, it would be impossible - there is not enough bandwidth in the PCH to begin to support what 12 TB3 links could demand, by an order of magnitude.
With the PCIe switches, all bets are off on slot configuration. I thought they said what they support when they introed it - maybe 4x16, 3x8, 1x4? Mike, do you remember?
I wouldn’t be surprised if I was wrong with the MCM packaging. It takes a rather large amount of time to keep track of Intel hardware, something I really don’t have time for. Intel also doesn’t make it easy with ARK being with obtuse. Really didn’t think they would put 64 PCIe lanes off the CPU just yet (on a SkyLake derivative), and dual die would be a cheap way for them to do it.
There should be about 84 PCIe lanes total with 64 off the CPU and 20 off the PCH. I don’t know what bandwidth or PCIe lane equivalent of the PCH connection to the CPU. If it is 8x, I can see those 20 used for the single width slots. If it is 4x, maybe the slot used for IO, 10G Ethernet ports, the TB controller for the top ports and the T2, and they’ll need a switch if they are advertising 16x support for the 7 full length slots.
If there is a switch, we will have to wait and see when someone decides to put say 6 Vega cards in it. There’s going to be a penalty.
64 lanes isn't all that surprising- the previous gen had 48, and AMD has been punishing them on this front for a couple of years already.
The lanes from the PCH are still equivalent in bandwidth to PCIe 3 x4, so are not terribly useful for TB (it has to support those 6k monitors, and just one of those will fill the entire PCH link). At best, it'll handle the two 10GbE, the two USB-A, and the internal SATA connectors. Hm, come to think of it, the two TB3 connectors on the IO card are limited (by that card's PCI link) to x4, so possibly that's going to the PCH. I doubt it though- I expect everything's routed through the switch(es).
You know, some of these questions are actually already answered on Apple's tech specs page. The slots are listed as x16, x8, x16, x16, x16, x8, x8, x4. So there is obviously at least one switch in use. Furthermore, there are two "Extra" PCI-ish slots in line with slots 1 and 3, which appear to be used only if slots 2 and 4 are not (though they are not specifically stated a being mutually exclusive, they likely are), which support x8 bandwidth each.
FWIW, even 7 Vegas seems easily doable: 4 in the the two MPX "bays", and one each in the other x16/x8 slots. Very few applications would notice even minor bandwidth limitations. You don't need this to drive six displays though - the dual-Vega MPXes can each drive three pro displays.
Apple tends to solder the processor to the board directly, rather than using any sort of holstering system, as a means to prevent processor changes after purchase, as well as potentially saving space by not needing the slotting mechanism.
The main reason to solder chips rather than socket them is the increase in reliability. Each contact in a socket adds a potential point of failure in the future.
Same for car wheels & bolts. Try to be reasonal.
He is being reasonable. Soldered components are more reliable than socketed. You're entitled to your own opinions, but not your own facts.
Due to thermal expansion, this is frequently not the case. For example, the Xbox 360's "red ring of death" was frequently caused by thermal expansion tearing the soldered processor off the board. Microsoft added better cooling and used epoxy to secure the corners of the chip to the board to attempt to address this.
LGA sockets have both more "give" and more mechanical support at the edges. That problem would not have happened with an LGA socket instead of a soldered processor.
an 8-core 3.5GHz processor ..., and supports up to 1TB of 2666MHz memory. 12-core chip is clocked at 3.3GHz, ... it also supports up to 1TB of memory the 16-core 3.2GHz ...it also includes support for 1TB of 2933MHz memory. the 24-core 2.7GHz ...the 24-core model can handle up to 2TB of 2933MHz memory, though Apple rates the Mac Pro to handle 1.5TB at this time. Lastly, the high-end 28-core processor is clocked at 2.5GHz, ...with the same memory capacity as the 24-core model.
So if some models are limited to 1TB and some 1.5TB, does that mean the max memory per slot is lower for the 1TB versions, or is the number of usable memory slots reduced for the 1TB versions?
The answer to this one is a little weird. You are likely to be able to use 128 GB LRDIMMs in any slot. Once you hit the 1 TB limit, it could go one of two ways. It is possible the limit is real, and additional sticks won't be used by the processor.
I think it is more likely the system only "supports" 1 TB, but will use anything you give it. "Support" in this context is purely about the ability to call the vendor for help if something goes wrong. Intel supports 32 GB of RAM with the NUC6 line of systems, but it has since been discovered you can use 64 GB (32 GB SO-DIMMs had not been shipped when those NUCs were sold).
As for the 2 TB limit, I strongly suspect that will not be reachable in the Mac Pro as shown. It would require 12x 170 GB DIMMs, which are not a thing. The limit on the processors is probably 2 TB for systems with 16 RAM slots.
Edited to add: Or you will have to use 4x256 + 8x128 sticks to reach it. That's a possibility, but would break bank interleaving, so the RAM would get significantly slower.
Would be neat if the Mac Pro shipped with the ability to take Optane DIMMs. It's easy to get to 2 TB with those.
The product page says up to 6 channel memory, but the 32GB comes with 4 sticks. I hope that's still with the same amount of slots and chipset, and not a lower end one with quad channel as LTT is suggesting.
There is absolutely no sign of movement in the Nvidia thing since we talked about it a few months ago.
Yes, but the important thing here is that is a OS/political issue; there is nothing in the hardware that prevent using Nvidia card if the two companies
One subject that hasn't been discussed a lot is storage; of course you can add PCI-e nvme cards (even raid), but i am talking about the builtin storage; the technical specification page talk about 2 SSD modules, and the drawing clearly show modules in the .m2 style, but it is not clear if they are standard or proprietary.
Have to wait and see on that. I would bet they are the same type of module Apple uses in the iMac Pro. The 2019 Mac Pro has a T2 coprocessor which contains the SSD controller. Apple’s ”SSD” daughter cards are basically dumb NAND only cards. Plus, they both look like 50 mm length cards, maybe 60 mm, that fit into slots of only a certain height. That looks limiting as far as how much storage can be put in there, and Apple will have to wait on density improvements to get more TB.
Overall, the storage story looks really good, and can look even better. Promise has a 4 3.5” drive MPX module, a 2 3.5” drive version, and obviously, you can get a regular PCIe SSD to slot into one of the slots, hopefully.
Comments
https://www.amazon.com/Cuisipro-Surface-Technology-4-Sided-Grater/dp/B003ZHU8M0?psc=1&SubscriptionId=AKIAILSHYYTFIVPWUY6Q&tag=duckduckgo-ipad-20&linkCode=xm2&camp=2025&creative=165953&creativeASIN=B003ZHU8M0
Later this year, both AMD and Intel will have some CPUs ready for 4, but not right now. From what I remember, Intel will have some Xeons. But it’s possible that it’s too late.
4 is just becoming available in some mobocracy. But even there it’s not all 4, but a hybrid of 3 and 4. 5 isn’t expected until 2021. The problem we’re seeing is that 3 has been around too long.
There should be about 84 PCIe lanes total with 64 off the CPU and 20 off the PCH. I don’t know what bandwidth or PCIe lane equivalent of the PCH connection to the CPU. If it is 8x, I can see those 20 used for the single width slots. If it is 4x, maybe the slot used for IO, 10G Ethernet ports, the TB controller for the top ports and the T2, and they’ll need a switch if they are advertising 16x support for the 7 full length slots.
If there is a switch, we will have to wait and see when someone decides to put say 6 Vega cards in it. There’s going to be a penalty.
LGA sockets have both more "give" and more mechanical support at the edges. That problem would not have happened with an LGA socket instead of a soldered processor.
The answer to this one is a little weird. You are likely to be able to use 128 GB LRDIMMs in any slot. Once you hit the 1 TB limit, it could go one of two ways. It is possible the limit is real, and additional sticks won't be used by the processor.
I think it is more likely the system only "supports" 1 TB, but will use anything you give it. "Support" in this context is purely about the ability to call the vendor for help if something goes wrong. Intel supports 32 GB of RAM with the NUC6 line of systems, but it has since been discovered you can use 64 GB (32 GB SO-DIMMs had not been shipped when those NUCs were sold).
As for the 2 TB limit, I strongly suspect that will not be reachable in the Mac Pro as shown. It would require 12x 170 GB DIMMs, which are not a thing. The limit on the processors is probably 2 TB for systems with 16 RAM slots.
Edited to add: Or you will have to use 4x256 + 8x128 sticks to reach it. That's a possibility, but would break bank interleaving, so the RAM would get significantly slower.
Would be neat if the Mac Pro shipped with the ability to take Optane DIMMs. It's easy to get to 2 TB with those.
Overall, the storage story looks really good, and can look even better. Promise has a 4 3.5” drive MPX module, a 2 3.5” drive version, and obviously, you can get a regular PCIe SSD to slot into one of the slots, hopefully.