nMP vs Workstation Price Comparisons

[iBug2]

I didn't really mean a niche within the existing market exactly; in fact, for all the joking about it being a Mac Mini Pro, the new Mac Pro does share a number of the Mac Mini's advantages thanks to being very compact and (sort of) easy to slot into an existing setup. It may also attract more of those people who are happy to stick with a high end iMac who are put off by the vast tower of old, i.e - those that simply wanted more power and didn't care about the rest of it.

Obviously it's ability to attract switchers would have been helped by having a much more affordable entry level model, so that's disappointing as they could easily have made an appealing $2,000 model from a high end Haswell i7 and a consumer graphics.

I think it'll sell quite well, at least as well as the previous Mac Pros, maybe even a bit better. Probably not to all the same people, but also to some people who wouldn't previously have considered a Mac Pro.

Anyway, this is a little bit off topic, my main point was that the Mac Pro has advantages and disadvantages that are difficult to put a price on

Indeed. I also think it'll sell along the lines of the old Mac Pro, but probably a bit less since like I said, every year less people need workstation class computers. That is inevitable.

To me, this is a much better deal than the old design. I hated the noise, I hated the size and weight. To me, the new one is a match made in heaven, except the dual GPU's, which I have no use for. But I'm ok with single processor setup, since 24 cores is overkill.

 

[fixmdude]

We can find out.

Only one way to know for sure about the speed of SSD drives in RAID compared to the nMP: I ordered the attached config from http://www.thinkmate.com
Xeon motherboards only have (2) SATA 3.0 6Gb/s ports on them, the other ports are all SATA 2.0 @ 3 Gb/s, so the only way to get enough speed is to add a RAID card. For this setup I selected a card with Cache onboard, which can be turned on and off for additional testing. As a result, the cost is higher than it needs to be for cost comparison, so this setup is just for speed test comparisons. It has 8 SSDs so we'll be able to see the various speeds as drives are added and find any cap, if there is one. I also selected a SATA 12Gb/s next generation controller which just came out. Even though the drives are 6Gb/s, the faster controller should be able to handle 8 drives without being its own bottleneck.

Advantage of the nMP:
- small
- light
- quiet
- power efficient
- 2 video cards
- cheap entry price for a Xeon workstation with ECC RAM
- fast, in a small footprint
- 6 thunderbolt expansion ports

Advantage of the ThinkMate:
- 64GB 1866MHz ECC RAM only fills 4 of 8 RAM slots, allowing for 128GB RAM (you have to ask for 1866MHz, their web site only shows 1600MHz due to availabilty but they will supply it upon request)
- 8 hot-swap drive bays which allows for flexibility and potentially faster speed if you spend more plus get a RAID PCIe card
- ThinkMate allows for the specs of the nMP without breaking the bank
- 6-core Xeon processor could be selected, saving ~$1600 over the 8-core Xeon now, with the option of upgrading to something newer/faster in the future
- 12Gb/s drives could be added in the future
- more drives allows for redundancy, RAID 10, unavailable on the nMP with 1 drive

Disadvantage to the ThinkMate:
- huge
- heavy (SuperMicro server chasis)
- only (2) PCIe 3.0 x16 slots which are filled with 1 video card and 1 RAID card so you can't have 2 video cards plus a x8 RAID card unless a video card fits into one of the other (2) x4 smaller slots
- the remaining 2 slots may be filled with a Thunderbolt card and a WiFi card, leaving fewer total ports/slots/expandability compared to the nMP

Disadvantage to the nMP:
- no internal drive redundancy option
- nMP doesn't exist, no availability date, important to know if money is to come out of 2013 budget

 

[linuxcooldude]

Nice try - but that's a comment about an early 2007 chipset designed for the first Core 2 Duo CPUs.

How about a test using a C602 chipset which might even be relevant?

You don't have a problem with pulling out the Mac Cube from 12 years ago with reference to the new Mac Pro.

 

[flat five]

If your apps do not use Open CL, that means you can simply go for D300 setup and pay around 250$ extra for a GPU you won't use. And yes, you can find other comparable setups which cost less (not a lot less) but are more suited for you.

just guessing..

it's possible the two gpus will divvy up tasks amongst themselves which, in turn, will make them last longer because of lighter usage..
in my experience, i burn up gpus more than any other component (approx every 2 years) so having two of them at the cheaper cost of buying as a packaged deal instead of separately might actually be a good thing?

 

[haravikk]

Only one way to know for sure about the speed of SSD drives in RAID compared to the nMP

You might want to test both RAID-0 and RAID-5 configurations; lots of people combine two SSDs with a RAID-0 for "double" the performance but, as with HDDs, it only really works for sequential access such as large files.

However, RAID-5 with three HDDs offers similar sequential read as the two drive RAID-0, while also allowing better potential speed for random access as well as random blocks can be fetched from potentially all three drives at once depending upon how they're distributed. It's not something I'd want to use SSDs for if the setup needed to do a lot of writing, due to the extra parity operations, but for read-heavy environments it's potentially better, plus it also gives redundancy.

It does generally require the RAID controller though, especially for OS X use.

 

[fixmdude]

You might want to test both RAID-0 and RAID-5 configurations; lots of people combine two SSDs with a RAID-0 for "double" the performance but, as with HDDs, it only really works for sequential access such as large files.

However, RAID-5 with three HDDs offers similar sequential read as the two drive RAID-0, while also allowing better potential speed for random access as well as random blocks can be fetched from potentially all three drives at once depending upon how they're distributed. It's not something I'd want to use SSDs for if the setup needed to do a lot of writing, due to the extra parity operations, but for read-heavy environments it's potentially better, plus it also gives redundancy.

It does generally require the RAID controller though, especially for OS X use.

From what I've found on a YouTube video created by an advertising firm hired by Samsung, (24) 840 Pro drives when in RAID0 can read at 2000MB/s. That's the same speed that others have gotten using RAID0 with only 4 of those drives, so the RAID cards must be getting saturated at 2000MB/s.

I went with 8 drives, which will allow RAID10 across 4 sets of 2, or, RAID50 across 2 sets of 4, for both speed and redundancy testing. With the new SATA 12 controller I found (SATA 4? 12Gb/s), I'll also be able to see if I can break the 2000MB/s barrier by using 4+ drives in RAID0.

But that's Sequential Access. The Random Access 4k question is a big one right now. Will that improve by RAIDing multiple SSDs or not? Some say it won't improve at all in RAID0. Some say it will. Maybe it will in RAID5.

 

[AidenShaw]

However, RAID-5 with three HDDs offers similar sequential read as the two drive RAID-0, while also allowing better potential speed for random access as well as random blocks can be fetched from potentially all three drives at once depending upon how they're distributed.

This just isn't true. The 3rd location contains parity information which is meaningless without data from the 1st or 2nd locations. (3rd + 1st => 2nd, 3rd + 2nd => 1st)

Edit: never mind - what I said is true for large or sequential reads - haravikk points out for small random reads you could do 3 at once....

 

[haravikk]

This just isn't true. The 3rd location contains parity information which is meaningless without data from the 1st or 2nd locations. (3rd + 1st => 2nd, 3rd + 2nd => 1st)

I'm talking about random/multi-threaded access, as RAID-5 uses distributed parity so the parity blocks are staggered, like so (each column is a separate disk):

A1  A2  Ap
Bp  B1  B2
C2  Cp  C1

So yeah, if you wanted to fetch the first three blocks (A1, A2, B1) then you're limited to only fetching from two drives at once. However, if you wanted to fetch the blocks A1, B1 and C1, then you can potentially fetch from all three drives simultaneously.

I'm not saying you'll get 3x performance, as it's pot luck whether you'll get the right distribution of blocks to ever sustain that, it's just one of those classic "up to" scenarios, but you definitely can get between one to three times the performance

 

[iBug2]

just guessing..

it's possible the two gpus will divvy up tasks amongst themselves which, in turn, will make them last longer because of lighter usage..
in my experience, i burn up gpus more than any other component (approx every 2 years) so having two of them at the cheaper cost of buying as a packaged deal instead of separately might actually be a good thing?

Good point but when you burn a GPU and buy a new one, you upgrade it as well. This way maybe you will be using the same GPU's for 5 years but they will feel pretty slow in the end as well.

 

[flat five]

Good point but when you burn a GPU and buy a new one, you upgrade it as well. This way maybe you will be using the same GPU's for 5 years but they will feel pretty slow in the end as well.

true.. that's how it goes for me.. I buy the cheaper replacements when one breaks but I'm still getting upgrades along the way..

makes me wonder about having mismatched GPUs in the nmp. as in, can someone have a d300 and assign it to display use only then a d700 which is assigned to computational tasks.. or if the OS optimizes in the background.. or if it's not possible at all..
we'll see

 

[haravikk]

But that's Sequential Access. The Random Access 4k question is a big one right now. Will that improve by RAIDing multiple SSDs or not? Some say it won't improve at all in RAID0. Some say it will. Maybe it will in RAID5.

It's the random access I'm interested by for comparing RAID-0 and RAID-5; although a 5 disk RAID-5 has almost identical sequential read performance to four disk RAID-0, the RAID-5 has potentially better random read performance (though probably similar or slightly worse random write performance) due to the way blocks are distributed.

This is the case with HDDs, and it should be the case with SSDs too, but there are lots of extra factors with them thanks to compression, the way blocks are aligned internally etc.

 

[fixmdude]

The initial speed test is in. Moving the PCIe 3.0 12Gb/s controller card from a x4 slot to a x16 slot so that it could use the card's full x8 speed increased the speed of the drives from 2,000 MB/s to 2,900 MB/s. That means there is only 1 x16 slot left open for a video card.

I'm surprised to see that the new Mac Pro is only getting around 900 MB/s, I would suspect that whoever ran the test is using the default of 64kb blocks instead of turning it up to 8 MB blocks to find out the drive's maximum possible output.

When mine is set to test 64kb blocks it drops down to 1,250MB/s

 

[slughead]

Good Grief... as has been stated numerous times in this thread... this thread is not about debating the merits of Apples design (there are at least 100 other threads already discussing that topic ad nauseum), it's merely about trying to price comparable systems.

You want it your way--find what the nMP is best at and try and find a PC that does the same at the same speed. What about finding a PC and trying to get the Mac to do the same. How is that invalid? How many pros are really going to be fine with 256GB of hard drive space?

Not everyone runs OpenCL and Disk Speed Tests for a living.

Anyways we don't know how the D700 actually does with professional tasks and hot benchmarks, so picking an appropriate GPU to pit against it is difficult.

----------

For many looking at the Mac Mini Pro, this quote from Jobs is relevant:

Amazing quote!