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

[danielwsmithee]

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.

My guess is that it is due to a bottle neck in the T2 chip. The two iMac Pro modules striped should theoretically give you theoretical read performance as high as 7000MB/s and write performance as high as 4200MB/s.

Now there is a bottleneck of the PCIeX4 interface of 4000MB/s. There is also going to be bottleneck in the T2 chip that receives reads/writes from the PCIeX4 bus and stripes them out to each drive, while additionally encrypting and decryption the data with its hardware AES encryption engine. My guess there is a read performance bottleneck in the T2, that is worse than the write performance bottleneck.

 

[Michael Scrip]

My guess is that it is due to a bottle neck in the T2 chip. The two iMac Pro modules striped should theoretically give you theoretical read performance as high as 7000MB/s and write performance as high as 4200MB/s.

Now there is a bottleneck of the PCIeX4 interface of 4000MB/s. There is also going to be bottleneck in the T2 chip that receives reads/writes from the PCIeX4 bus and stripes them out to each drive, while additionally encrypting and decryption the data with its hardware AES encryption engine. My guess there is a read performance bottleneck in the T2, that is worse than the write performance bottleneck.

That makes sense.

Thank you, and @CWallace

 

[leman]

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

Does the 960 Pro actually reach those speeds in tests? All actual benchmarks I saw show it in the ballpark of ~2000MB/s, give or take. I think we should wait for some proper benchmarking of the iMac Pro storage system before discussing this further.

 

[danielwsmithee]

Does the 960 Pro actually reach those speeds in tests? All actual benchmarks I saw show it in the ballpark of ~2000MB/s, give or take. I think we should wait for some proper benchmarking of the iMac Pro storage system before discussing this further.

This is a good point. We've been talking about the advertised speed of the 960 Pro that Samsung uses for marketing and labeling. My guess is it is theoretical and only achievable under the best circumstances.

According to this:
http://ssd.userbenchmark.com/SpeedTest/182182/Samsung-SSD-960-PRO-512GB
The performance is closer to 2200MB/s read 1700MB/s write.

Here are some interesting results of a RAID0 combination of two M.2 NVME SSDs compared to their individual drive performance.
https://www.eteknix.com/year-nvme-raid-0-real-world-setup/6/

You will notice significantly different results based on which benchmark is being RUN. Some (IOPS) have the read speed actually decreasing while the write speed increases. My guess is this is what is happening with Blackmagic on the iMac Pro. Other benchmarks may provide very different results.

It looks like striping two NVME SSDs increases read latency, but does not impact write latency. Which makes sense given that the T2 is going to have reassemble the stream of READ data as it comes off each drive, while writes it just needs to split out between the drives.

 

[chucker23n1]

Someone who dug deeper feel free to correct me, but it looks to me like Blackmagic writes a file
DiskSpeedTestTemp (on your regular file system!*), then reads it back. IOW, its measurement are already net values with the overhead of Apple's controllers, Apple's encryption, the APFS file system and other simultaneous I/O subtracted.

Samsung's given specs are very different values than that.

*) in a sandboxed container in your user directory:
~/Library/Containers/com.blackmagic-design.DiskSpeedTest/Data/Library/Caches/BlackmagicDiskSpeedTest/DiskSpeedTestTemp

 

[Wildhope]

Ugh this makes me want one. I can never afford it and way too much power for me. lol

 

[DesterWallaboo]

Apparently BareFeats has some new benchmarks posted today. They include some After Effects benching. Surprisingly the iMac Pro was faster than the iMac 5K at renders by ~30%. I've requested the AE project from Rob at BareFeats to see if the project is using purely AE standard plugins and features, or if the project is using any common 3rd party plugins (Red Giant, Trapcode, Video CoPilot, etc.).

Though I've not seen the project (yet), I'm guessing the improved benchmarks are primarily due to the project using only native plugins and features where Metal acceleration under AE CC 2018 would be massively benefitted with the AMD Vega card. As of yet, Red Giant/Trapcode and Video Copilot are not using full Metal acceleration, though they do use OpenGL GPU acceleration. I also believe they use some CUDA, though I'm not certain on that one. Of course CUDA wouldn't be beneficial for an AMD card.

In our experience, with past iterations of After Effects without Metal acceleration, CPU clock speed trounces multiple cores and HT.

I should add that our projects almost never use purely After Effects base plugins and features. So testing projects that include Red Giant/Trapcode and Video Copilot (amongst others), would be a much better method for testing how well the iMac Pro performs compared to higher clock speed Macs.

http://barefeats.com/imacpro_vs_pt2.html

 

[leman]

would be a much better method for testing how well the iMac Pro performs compared to higher clock speed Macs

The iMac is not higher clocked. Don’t let the vase frequency mislead you. The top clock of these Xeons (except the 18-core part which is slightly slower) is 4.5ghz, which is identical to the highest-performance iMac CPU

 

[DesterWallaboo]

The iMac is not higher clocked. Don’t let the vase frequency mislead you. The top clock of these Xeons (except the 18-core part which is slightly slower) is 4.5ghz, which is identical to the highest-performance iMac CPU

I can tell you in our testing with Mac Pros and nMP's that the 5K iMacs are significantly faster than the Mac Pros or nMP's under the same tasks on AE. Also, from some of the tests I saw, it appears that the iMac Pro tends to wait until the CPU gets pretty hot before cranking up the fans, and it chooses to instead drop the clock speed instead of cranking up the fans. I expect either there will be changes in the default OS settings for the fans, or that 3rd party implementations will resolves this.

 

[danielwsmithee]

I can tell you in our testing with Mac Pros and nMP's that the 5K iMacs are significantly faster than the Mac Pros or nMP's under the same tasks on AE.

The iMac Pro is significantly higher clocked than the Mac Pro or nMP. I would expect the iMac Pro to have slightly better single core performance on most tasks than the 5K iMac and significantly better multi-core performance.

 

[DesterWallaboo]

But the golden question remains.... is the ~30% increase under After Affects worth the the extra $2,000+ of the iMac Pro? We're talking a nearly 60% higher purchase cost for the basic model. Which, having looked at configurations, wouldn't really meet our needs. We would probably be sitting closer to $6,000 or $7,000 for what we need here.

 

[danielwsmithee]

But the golden question remains.... is the ~30% increase under After Affects worth the the extra $2,000+ of the iMac Pro? We're talking a nearly 60% higher purchase cost for the basic model. Which, having looked at configurations, wouldn't really meet our needs. We would probably be sitting closer to $6,000 or $7,000 for what we need here.

I doubt it is worth the cost differential especially when you consider the 5K iMac will likely be updated to to the i7-8700k next year which turbo boosts higher to 4.7 and has 6-cores.

 

[phobos]

A silent working environment - to me at least - is worth some extra money. It seems the iMac Pro is very quiet which is the exact opposite of the regular iMac.

For example when you’re trying to set up the lighting and materials of a 3D scene you have your computer constantly rendering, which means the CPU or GPU are at 100%. Within seconds the regular iMac has the fans spin up at really high speeds. Having to work in this constant noise would just be terrible, and in the end your work would suffer because you’re compromising just to keep the machine silent. Not to mention being afraid to leave the machine on overnight in fear of it melting!

Even if you’re not constantly stressing the CPU and GPU I think the iMac Pro’s internals will have a longer lifetime, just from the better thermals alone.

 

[JoeG4]

Oh, and no anti static wrist strap, WTF, or do they throw this one away.

If you're in a lab a lot you can get a shoe strap instead of a wrist strap - the blue pad they're working on would usually be grounded too. That said, it looks like they did this video in a conference room. LOL. They're likely not even trying.

 

[justperry]

If you're in a lab a lot you can get a shoe strap instead of a wrist strap - the blue pad they're working on would usually be grounded too. That said, it looks like they did this video in a conference room. LOL. They're likely not even trying.

We have an electronics assembly lab in my company, so I know a thing or 2 about it, the floor is ESD safe and they all wear shoe straps, before they enter the lab they have to test the straps.
I myself also used shoe and wrist straps, a couple of years ago someone had to fit boards in an electronic device, out of those 20-30 boards about 4 were damaged cause the guy didn't wear a strap.

 

[leman]

But the golden question remains.... is the ~30% increase under After Affects worth the the extra $2,000+ of the iMac Pro? We're talking a nearly 60% higher purchase cost for the basic model. Which, having looked at configurations, wouldn't really meet our needs. We would probably be sitting closer to $6,000 or $7,000 for what we need here.

I guess it depends how much money you are making out of it. I would assume that the revenue generated from a single workstation far exceeds the computer costs. I mean, you will most likely make more then 8k from a machine per three years And if 30% performance increase allows you to increase your revenue by 10% or so, then of course it’s worth it.

 

[bplein]

To everyone who thinks that Apple running RAID 0 is bad/dangerous: Read what leman said: Inside ~any~ SSD is are several chips. The SSD stripes (RAID 0) across these several chips.

These systems have a single logical SSD as seen from the OS. Your job is to back it up so that if a part does fail, you don't lose your data and can recover.

RAID 0 does double the risk of failure. But what is that risk? VERY LOW.
[doublepost=1514650059][/doublepost]

Yes I’m aware of that. That’s exactly what I said on my post as well.
Just in case you’re not aware RAID0 means that if something goes wrong with one drive you lose both drives.

No you don't lose both drives. You lose your data on both drives. When one drive failes, the other is perfectly healthy. When you get your iMac Pro repaired, they will replace ONE DRIVE not two.

Please, stop it with this nonsense.

 

[phobos]

No you don't lose both drives. You lose your data on both drives. When one drive failes, the other is perfectly healthy. When you get your iMac Pro repaired, they will replace ONE DRIVE not two.

Please, stop it with this nonsense.

I think you just want to play with words. I’m pretty sure you understood what I said. You lose data on both drives and not the one drive. Because that's how RAID0 works. And that’s the whole point.
Just because you don’t like where the conversation is going doesn’t mean that people need to stop discussing about it.

 

[bplein]

I think you just want to play with words. I’m pretty sure you understood what I said. You lose data on both drives and not the one drive. Because that's how RAID0 works. And that’s the whole point.
Just because you don’t like where the conversation is going doesn’t mean that people need to stop discussing about it.

Words have meanings. If you meant data, say data.

If you are afraid of RAID 0, then OK. As a storage professional working specifically in the area of enterprise grade PCIe Flash, SSD-based storage arrays, and NVMe storage for over 8 years, I can tell you that your fears are overblown, that you need to backup your data regardless of RAID or JBOD levels of storage, and that everything you've said in this thread could be altered to make a single SSD look stupid and dangerous in comparison to using RAID 1. But you aren't saying that everyone should mirror their boot drives.

Furthermore, you (and I as well) know ZERO about the expected longevity of these devices, either from an MTBF perspective or from a write endurance perspective. 2-drive RAID zero DOES double the chance of a single failure. But not knowing the MTBF means you cannot compare the risk of this solution to any other known solution. Using 2 devices also allows for double the garbage collection speeds (something that throughput testing ignores and cannot estimate).

There are lots of things we don't know about this, all we know is what we can see from early Apple photos (which may have been computer generated anyhow) and the one tear down that we have access to so far.

The rest is speculation. Like your first impressions video you released this summer.

 

[leman]

I think you just want to play with words. I’m pretty sure you understood what I said. You lose data on both drives and not the one drive. Because that's how RAID0 works. And that’s the whole point.
Just because you don’t like where the conversation is going doesn’t mean that people need to stop discussing about it.

Actually, the one playing with words is you. You’ve been informed, by multiple people, that every single SSD on the market right now is using a RAID-0-like setup (something that you still haven’t acknowledged explicitly or replied to btw). If striped configurations worry you that much, then you shouldn’t use SSDs in the first place. Frankly, there is no real need in reiterating urban myths that long became obsolete.

More objectively, is using two 512gb SSDs inherently less reliable than using a single 1TB SSD? Maybe. This depends on multiple factors such as: how many modules of which densities do these SSDs use internally? Is there a difference in actual module reliability? What are the environmental factors? Your „it’s striped, so you are much more likely to loose your data“ attitude is a bit simplistic. I am sure that Apple did extensive testing here and that they are confident enough to support this use case with warranty. Don’t forget that repairs are rather expensive for Apple, and given the relatively low prices they sell these units at, high amount of failures will quickly make them lose money on sales.

 

[tipoo]

Since there's one controller on the T2 that's just handling two physically separate banks of NAND, I don't think it makes sense to call this a RAID. That would be a T2/controller on each NAND bank. This is just one controller looking in two places and unifying the NAND away from the OS, as it would for multiple chips, just more physically separated now.

 

[phobos]

Furthermore, you (and I as well) know ZERO about the expected longevity of these devices, either from an MTBF perspective or from a write endurance perspective. 2-drive RAID zero DOES double the chance of a single failure. But not knowing the MTBF means you cannot compare the risk of this solution to any other known solution. Using 2 devices also allows for double the garbage collection speeds (something that throughput testing ignores and cannot estimate).

There are lots of things we don't know about this, all we know is what we can see from early Apple photos (which may have been computer generated anyhow) and the one tear down that we have access to so far.

The rest is speculation. Like your first impressions video you released this summer.

That's all I'm saying. You're unecessarily increasing your chances of failure.
For my RAID system I have dual disk redundancies setup. If also a drive fails I can still work with my files and get on with my work. No real downtime. I can pop in a new hard drive and as the system rebuilds the drive I can keep working.

With the RAID 0 setup for a startup drive if something happens I'm completely incapacitated until everything is setup again. It's the computers drive. I need to stop whatever work I'm doing, push further whatever deadlines I have and deal with the issue. Because I have no other choice. So in this case yes I do have everything backed up, but that's not really going to help me with the downtime and the long time I need to setup every little thing as I want because not everything is saved on the Time machine backups.

I feel that for a system's boot drive things should be as safe as possible.

Either way we'll just have to wait and see. I certainly hope the drives are as solid as we want them to be. I definitely don't want to deal with something like that on a really busy day.

Actually, the one playing with words is you. You’ve been informed, by multiple people, that every single SSD on the market right now is using a RAID-0-like setup (something that you still haven’t acknowledged explicitly or replied to btw).

I am sure that Apple did extensive testing here and that they are confident enough to support this use case with warranty. Don’t forget that repairs are rather expensive for Apple, and given the relatively low prices they sell these units at, high amount of failures will quickly make them lose money on sales.

First of all I'm going to disregard your tone. I was responding to bplein and not you. And I'm not the only one in this thread raising their concerns. So I don't see why you feel the need to personally address me. I understand how SSDs work. I also know that having a lot of different points of failure, you're increasing your chances of failure.
It's as simple as that.

And as far as extensive testing goes. As much as I like Apple I wouldn't put so much trust to it. I don't think I need to remind anyone the overheating GPUs on the Mac Pro that would either destroy the cards or the motherboard, the blown Nvidia cards on MacBook Pros the broken hinges on the 2012 iMacs, the detaching backs on the Apple watches etc. The list is long and as you can see a lot of things can go through testing without noticing. And knowing the drop in quality the past few years I wouldn't be surprised if we saw some hardware problem with the new iMac Pro.

FYI I have already ordered one but it doesn't mean that I won't question any of Apple's decisions and take everything at face value. Apple is not an all knowing being that cannot make mistakes.

I don't really want to drag along the discussion. So if you want to, feel free to reply of course, I'm just not going to address this issue anymore

 

[leman]

First of all I'm going to disregard your tone. I was responding to bplein and not you.

Given the fact that we have directly communicated on this very issue in this thread, I felt entitled to answer your post

It's the computers drive. I need to stop whatever work I'm doing, push further whatever deadlines I have and deal with the issue. Because I have no other choice.

This sounds to me less like a criticism of the storage configuration Apple chose to use and more a criticism of the lack of user-serviceability. And of course, you point makes perfect sense. Apple's choice is clearly maximising performance over minimising downtime. It it worth it? No idea. Again, my personal opinion is that given the very low SSD failure rates, it won't matter much in practice (except the few less lucky users who get hit by failures), while increased speed will probably matter more to the overall user base.

 

[CWallace]

That's all I'm saying. You're unecessarily increasing your chances of failure.

It's not "unnecessary" because there is a clear performance benefit to using this configuration over a single-drive configuration. And even if you double you're failure rate (and you very likely are not), based on current SSD annual failure rates, we're talking a four-tenths of one percent chance of failure compared to a two-tenths of one percent chance which to me is still pretty good odds (especially with full backups).

 

[bplein]

I don't really want to drag along the discussion. So if you want to, feel free to reply of course, I'm just not going to address this issue anymore

I think my tone in my first response got us off in the wrong direction. I was mostly responding to everyone (in the world!) who say a lot of things that are based on their anecdotal experiences or based on "what they read", and I got fed up with it.

Some days I should just not hit "reply"