iMac Pro Base Model Teardown Reveals 2x SSD RAID Configuration and Four 8GB DIMM Modules

[maflynn]

Lots of RAID 0 comments which is interesting. Personally, I can see why they went for this arrangement as they wanted to get the best performance

Personally, I'm not a fan of it, and we really don't know if its RAID 0 or just their Fusion implementation (I've not read through the entire thread so I could be wrong and it is RAID 0). To me, it seems they're cheaping out, i.e., cheaper to get two 512GB SSDs vs. 1 TB SSD. How much speed increase will there be on a RAID 0 SSD implementation, there is no drive head latency, that is RAID 0 really was useful with hard drives.

 

[leman]

How much speed increase will there be on a RAID 0 SSD implementation, there is no drive head latency, that is RAID 0 really was useful with hard drives.

You can split the writes to multiple modules in parallel, effectively doubling the bandwidth. I don't expect there to be any change to latency. As others have mentioned before, the system sees the configuration as a single storage unit, meaning that its not a "software" solution. In essence, they probably just put another controller in front of two SSD controllers.

I agree with you that this is a way to increase performance and (probably) reduce the costs, compared to more traditional implementation. I do not know what SSD module configurations are currently used, but assuming that a 512GB and 1TB SSD use the same amount of storage modules, the 1TB version would need to use more expensive denser modules. Apple opts instead for double the amount of lower-density ons. Can they really save costs with it though? I wouldn't be so sure. After all, the investment in an additional RAID controller etc. is probably significant. Then again, they had to develop this thing in order to have 4TB SSDs in the first place.

At any rate, the performance of these SSDs rival the best enterprise stuff out there, and we are talking about modules that cost ~$2000 for 2TB for the retain customer and occupy a full PCI-E slot. I think its rather impressive what we have here, even though many people — as usual — take their time to understand what is actually going on.

 

[vmistery]

Personally, I'm not a fan of it, and we really don't know if its RAID 0 or just their Fusion implementation (I've not read through the entire thread so I could be wrong and it is RAID 0). To me, it seems they're cheaping out, i.e., cheaper to get two 512GB SSDs vs. 1 TB SSD. How much speed increase will there be on a RAID 0 SSD implementation, there is no drive head latency, that is RAID 0 really was useful with hard drives.

Could be a limitation on the drive controller itself? I dont know just guessing but when you look at most M.4 drives they don't seem to touc

Personally, I'm not a fan of it, and we really don't know if its RAID 0 or just their Fusion implementation (I've not read through the entire thread so I could be wrong and it is RAID 0). To me, it seems they're cheaping out, i.e., cheaper to get two 512GB SSDs vs. 1 TB SSD. How much speed increase will there be on a RAID 0 SSD implementation, there is no drive head latency, that is RAID 0 really was useful with hard drives.

According to the article they are two physically different modules held down with screws. Could be something like an interface limitation they are working around or that 2 * 512GB is substantially faster in some way than 1 * 1TB. I don't know, just guesses. As for RAID 0 on spinning rust I'd have spent the extra on gone for a RAID 1+0 as I don't trust disks to last (especially the 15k ones...)

 

[germinator]

Has anybody figured out what kind of SSD thise are?

 

[danielwsmithee]

Whether the T2 Controller implements a RAID0 to stripe the drives together or concatenates the volume the end result is more or less the same, especially if encryption is enabled.

The interesting thing is NVME volumes top out at 3500 MB/s, and 2000MB/s write performance. So a RAID0 could theoretically top out at 7000MB/s and 4000 MB/s. I haven’t seen any write performance benchmarks yet on the iMac Pro. I suspect that if we are seeing 3000MB/s read performance the performance is limited by the T2 Encryption engine anyways and the whole things sits on a single PCIex4 Interface. Whether they are spanned together or striped we could likely tell by the write performance, but the end result is the same.

I for one do hope that the new Mac Pro has 2 or 3 standard M.2 slots and supports RAID configurations of those slots.

I just found this:

Part of the huge performance boost stems from the speed of the iMac Pro’s SSD. According to Laforet, it managed a hefty write speed of 2,996MB per second, and a read speed of 2,450MB per second. Meanwhile, the MacBook Pro’s 1TB SSD saw a write speed of 1,743MB/s and a read speed of 1,400MB/s.​

From https://www.digitaltrends.com/computing/imac-pro-early-reviews-arrive/

Based on those numbers my guess is the iMac Pro uses the same NVME modules as the MacBook Pro (though those are soldered down) and the T2 controller stripes data across the two modules, but it is presented to the system as a single NVME drive and only uses x4 PCIe interface.

 

[itguy06]

The Rizen 7 1700 averages 3163 in single-core and 18814 in multi-core.

The Xeon W-2140B averages 4715 in single-core (49% more) and 29080 in multi-core (55% more).

That's… stretching "about as fast" quite a bit.

Heck, you can configure the iMac non-Pro with the 7700K, and still end up well ahead of the Ryzen: 5613 single-core (77% more) and 21181 multi-core (26% more).

In which benchmarks? Geekbench? What exactly do those #'s mean? Is it linear, is it 1 for 1, what? Geekbench is quite coy about what their benchmarks actually mean other than lots of marketing speak.

Would be more interesting to see actual real world benchmarks like 4k editing, photo editing, gaming, etc. rather than some useless #s from a software that doesn't tell how they get those #s.

As for the 7700k it's a good CPU but for multicore workloads (which we're slowly moving to) it's behind the 1700. http://cpu.userbenchmark.com/Compare/AMD-Ryzen-7-1700-vs-Intel-Core-i7-7700K/3917vs3647 It's mainly due to its higher frequency (base and turbo) that it does so well. It also consumes more power and puts off more heat if that matters to you.

 

[CWallace]

To me, it seems they're cheaping out, i.e., cheaper to get two 512GB SSDs vs. 1 TB SSD.

Based on retail prices, a single 1TB m.2 SSD is cheaper than two 512GB models. Plus there is the additional cost of the second connector. So I doubt Apple went with the dual route to save money.

And are m.2 drives even available in capacities higher than 2TB? I believe not, so Apple would have to use two 2TB modules to reach the 4TB configuration anyway.

 

[Zdigital2015]

Has anybody figured out what kind of SSD thise are?

Dollars to doughnuts the SSD is a custom Samsung flash package. I would not doubt that it is based on the 960 EVO series, as Apple has OEM'ed other Samsung products have been used throughout Apple's product line for the past 6-7 years.
[doublepost=1514566982][/doublepost]

Based on retail prices, a single 1TB m.2 SSD is cheaper than two 512GB models. Plus there is the additional cost of the second connector. So I doubt Apple went with the dual route to save money.

And are m.2 drives even available in capacities higher than 2TB? I believe not, so Apple would have to use two 2TB modules to reach the 4TB configuration anyway.

You're 100% correct! Samsung offers the 960 EVO m.2 up to 2TB, but you have to go to the 2.5" form factor to get to 4TB. I'm convinced it has to do with heat dissipation as this is becoming more of an issue now with high density m.2 blades. Motherboard manufacturers have even jumped on the bandwagon with their own m.2 "armor" to help with thermal issues - RE: ASUS ROG STRIX Z370-E Gaming motherboard.

 

[mikeboss]

nope, the SSDs are Apple's own design. see here -> #4

 

[chucker23n1]

In which benchmarks? Geekbench?

Yes, those are Geekbench numbers.

What exactly do those #'s mean? Is it linear, is it 1 for 1, what? Geekbench is quite coy about what their benchmarks actually mean other than lots of marketing speak.

Would be more interesting to see actual real world benchmarks like 4k editing, photo editing, gaming, etc. rather than some useless #s from a software that doesn't tell how they get those #s.

No, dude. You made the claim that a CPU that costs a tenth is "about as fast". In benchmarks, it clearly is not. The onus is on you to prove your claim.

As for the 7700k it's a good CPU but for multicore workloads (which we're slowly moving to) it's behind the 1700. http://cpu.userbenchmark.com/Compare/AMD-Ryzen-7-1700-vs-Intel-Core-i7-7700K/3917vs3647 It's mainly due to its higher frequency (base and turbo) that it does so well.

Yes, and it's mainly due to having, oh, I dunno, twice the cores that the Rizen competes well against the 7700k in multicore.

Anyway, I couldn't find any Xeon-W CPU option on that site, and from the 1990s' GIFs and multiple mentions of CPU overclocking, I'm having a hard time taking it seriously. But it says the Rizen is a "Nuclear submarine", so I guess that's hilarious and cool.

It also consumes more power and puts off more heat if that matters to you.

My guess is that concern is secondary to the intended iMac Pro audience.
[doublepost=1514569146][/doublepost]

nope, the SSDs are Apple's own design. see here -> #4

That post literally shows a photo with a Samsung copyright.

So, yes, the SSD may be a custom Apple design (and the controller clearly is), but the flash chips at the very least are Samsung's.

 

[mikeboss]

That post literally shows a photo with a Samsung copyright.

So, yes, the SSD may be a custom Apple design (and the controller clearly is), but the flash chips at the very least are Samsung's.

the SSD with a SAMSUNG© on the right is the old SSUBX model.
the one found in the iMac Pro has a proprietary connector and is Apple's own design.

 

[chucker23n1]

the SSD with a SAMSUNG© on the right is the old SSUBX model.
the one found in the iMac Pro has a proprietary connector and is Apple's own design.

Hm, misread your post. But the first photo doesn't actually show the chips.

 

[leman]

You're 100% correct! Samsung offers the 960 EVO m.2 up to 2TB, but you have to go to the 2.5" form factor to get to 4TB. I'm convinced it has to do with heat dissipation as this is becoming more of an issue now with high density m.2 blades. Motherboard manufacturers have even jumped on the bandwagon with their own m.2 "armor" to help with thermal issues - RE: ASUS ROG STRIX Z370-E Gaming motherboard.

Another problem is the SSD module capacity. I guess there are simply no SSD modules that would allow one to fit more then 2TB enough on a standard m.2 part. I am quite sure that if you open up one of the 2.5" higher-density SSDs, you'll find a solution not unlike what Apple is doing in the iMac Pro

 

[CWallace]

nope, the SSDs are Apple's own design. see here -> #4

(Edited)

The connector is (probably) proprietary to Apple - I am guessing to interface with the T2 - as the power connectors are identical between the two modules, but the data connector is now 23 pins instead of 16.

However as noted by chucker23n1, the components themselves are most likely Samsung (controller and NAND) as Apple doesn't have an SSD manufacturing capacity and Samsung provides all their other SSD modules.

So this is likely a custom module built by Samsung to Apple's specifications and using Samsung's reference NVMe design.

 

[mikeboss]

The connector is (probably) proprietary to Apple - I am guessing to interface with the T2 - however as noted by chucker23n1, the components themselves are Samsung (controller and NAND). So this is likely a custom module built by Samsung to Apple's specifications and using Samsung's reference NVMe design.

as long as nobody disassembles one of these new SSD modules, we won't know for sure. but on the SAMSUNG built modules, there always was a sticker mentioning SAMSUNG. starting with the MacBook Retina, Apple began to develop their own SSDs. not the NAND, but the controller. and since there's a completely unknown connector on these new modules, I wouldn't be surprised if they were developed inhouse at Apple.

 

[danielwsmithee]

as long as nobody disassembles one of these new SSD modules, we won't know for sure. but on the SAMSUNG built modules, there always was a sticker mentioning SAMSUNG. starting with the MacBook Retina, Apple began to develop their own SSDs. not the NAND, but the controller. and since there's a completely unknown connector on these new modules, I wouldn't be surprised if they were developed inhouse at Apple.

According to iFixIt the new MacBook Pro uses SAMSUNG NAND with an Apple designed custom controller.

https://www.ifixit.com/Teardown/MacBook+Pro+15-Inch+Touch+Bar+Teardown/73395#s148876

I'm pretty certain at this point that each of these modules in the iMac Pro contains this same controller and NAND Flash chips and the T2 communicates with each of them and stripes the data across each. To reach the nearly double NVME performance of the MacBook Pro.

 

[TofuTodd]

arrived today

 

[Michael Scrip]

arrived today

That's fast... for sure.

But don't single MVNe drives hit those kinds of speeds?

I'm just wondering what Apple is gaining by RAIDing two drives together.

 

[CWallace]

as long as nobody disassembles one of these new SSD modules, we won't know for sure. but on the SAMSUNG built modules, there always was a sticker mentioning SAMSUNG.

One of the OWC teardown pictures you linked to show just that:

The MZ-KPV1T00/0A4 is the same Samsung 1TB SSD module used in the 2014-2017 5K iMac and 2013-2015 MacBook Pro. So this connector is actually used by all Apple Mac products and is not unique to the iMac Pro.

 

[mikeboss]

That's fast... for sure.

But don't single MVNe drives hit those kinds of speeds?

I'm just wondering what Apple is gaining by RAIDing two drives together.

indeed, modern NVMe SSDs also deliver this kind of speed.
problem is: the T2 connects to the PCH with only 4 PCIe 3.0 lanes.
[doublepost=1514573999][/doublepost]

One of the OWC teardown pictures you linked to show just that:



The MZ-KPV1T00/0A4 is the same Samsung 1TB SSD module used in the 2014-2017 5K iMac and 2013-2015 MacBook Pro. So this connector is actually used by all Apple Mac products and is not unique to the iMac Pro.

yeah, that's not from the iMac Pro. this is an old SSUBX module. I posted this just for comparison. I tried to edit my post to clarify that, but the forum software is not allowing that anymore.

 

[CWallace]

But don't single MVNe drives hit those kinds of speeds?

I have the 1TB SSD in the 2017 iMac 5K and while I get similar read speeds (2586MB/s), my write speeds are much slower (1785MB/s).
[doublepost=1514574121][/doublepost]

yeah, that's not from the iMac Pro. this is an old SSUBX module. I posted this just for comparison. I tried to edit my post to clarify that, but the forum software is not allowing that anymore.

Ah, ok. Was kind of wondering why it was 1TB module with the machine said to have two 512GB units.

The power connectors are identical between the two modules, but the data connector is now 23 pins instead of 16. I assume the extra pins are for the T2 connectivity.

 

[Rockadile]

Can the SSD be changed to RAID 1 or as independents?

 

[Michael Scrip]

I have the 1TB SSD in the 2017 iMac 5K and while I get similar read speeds (2586MB/s), my write speeds are much slower (1785MB/s).

Ohhh yeah. I just noticed the bigger number was the write speeds on the iMac Pro drives. Holy smokes.

Usually read speeds are much faster than writes on standard PCIe MVNe drives like the Samsung 960 EVO/PRO... somewhere in the neighborhood of ~3000MB/s read and ~2000MB/s write.

But I guess if these two Apple drives are in RAID0... the write speeds would be faster (since you're striping the writes)

Then again... wouldn't the read speeds in RAID0 also be faster than a single drive?

This is why I'm confused: the screenshot posted above shows the iMac Pro drives with 3011MB/s write and 2469MB/s read. In other words... faster write than read. And we're assuming these two drives are in RAID0.

But a single Samsung 960 EVO gets the opposite numbers... faster read than write.

 

[CWallace]

Can the SSD be changed to RAID 1 or as independents?

No. The T2 controller presents them as a single logical drive to macOS and you cannot change that in Terminal or Disk Utility.

But I guess if these two Apple drives are in RAID0... the write speeds would be faster (since you're striping the writes). Then again... wouldn't the read speeds in RAID0 also be faster than a single drive?

The SSD controller is connected via 4-lane PCIe 3.0 and the bandwidth limit for that configuration is 4000MB/s. Adding in controller overhead, and the peak speeds we are seeing represent the usable limit of that configuration.*

* - (This is not an Apple limitation. Most, if not all, Intel workstation and desktop chipsets use a 4x PCIe 3.0 link for the SSD module).

 

[Michael Scrip]

The SSD controller is connected via 4-lane PCIe 3.0 and the bandwidth limit for that configuration is 4000MB/s.

I still don't understand why Apple's writes are faster than their reads... when other single drives are the opposite.

In other words... why does the iMac Pro drive only get 2400MB/s read while a Samsung 960 Pro can get 3400MB/s read?

The Samsung 960 Pro is also using the 4-line PCIe bus.