What computer should I buy to make MATLAB simulations run fast?

I use MATLAB to run large population models. I represent populations as matrices which may have as many as 4,000 rows and up to hundreds of columns. Some simulations may track up to hundreds of matrices. I often subject the matrices to basic math operations in which entries are treated as scalars (e.g., exp(A), A.^2, etc.), or to commands such as sum, cumsum, max, flipdim, and sortrow.

I have the opportunity to buy a new computer on which to run these simulations, and I want to buy something that will run my simulations as quickly as possible. My organization is committed to the Mac platform. I would like to keep the total price in the US$4,000 to $8,000 range (after an institutional discount), although I could go slightly over if there were good reason to do so. On the other hand, I don’t want to throw money away on hardware that won’t make a meaningful difference in simulation speed.

The IT office at my institution has recommended:

Mac Pro

     12-core with 2x 6-core Intel Xeon 2.4GHz processors
     16GB RAM 
     ATI Radeon HD 5770 – dual output
    (~$3,400 base price; upgrade to 2.66GHz +$1,000; 3.06GHz +$2,000)
2x 1TB drives (to change to 0.5GB Solid State +$800 per drive)
2x Dell UltraSharp U2711 69cm(27") monitors (~$1,400)

Does this seem reasonable? Is it worth the extra money to get the faster processors? Are there things I should cut because they won’t help very much, or add because they will?

Answer 1

[Walter Roberson]

If you are looking at the Mac Pro, see if you can wait several months. The Mac Pro has not been refreshed for the Sandy Bridge architecture, or for the higher speed base graphics. There is also no official support on the Mac Pro for CUDA, and putting in a Tesla apparently requires hardware mods, it has been written here.

Apple did specifically indicate earlier this year that they are working on a Mac Pro refresh, but did not give a time frame.

When I was trying to decide what to buy earlier this year, I looked at benchmarks. They are, as always, somewhat confusing. The new MacBook Pro (laptop!) was nearly as fast for general work, and was faster for disk work if you got the SSD version, and was faster for some graphics. The Mac Pro only really shone in total computation power for tasks that could use the full 12 cores.

I do not know much about programming with CUDA, but what you describe your tasks as seems likely to me to be able to program relatively easily with CUDA; if that is correct it might change the balance of your decision.

[Jason Ross]

I completely agree with this, the Mac Pro is definitely "behind the times" as it's not taking advantage of the Sandy Bridge architecture like the remainder of the Mac line has already.

CUDA support does exist for Mac, but on the Mac Pro the only game in town is the Quadro 4000 for Mac. No Tesla (C2075, C1060) card will work with it out of the box. nVidia has announced an upgraded card, though, the K5000 (<http://www.zdnet.com/nvidia-outs-quadro-k5000-graphic-card-for-apples-mac-pro-7000004174/>) that should run in a Mac Pro. And there are people using the GPU in the MacBook Pro, as well -- although that GPU also doesn't have the same processor count and RAM as a Tesla card.

I'm also a big fan of stuffing as much RAM into a machine as possible.

The SSD is the wave of the future, allowing quick access, no moving parts and lower power consumption. Only you can decide if you are ready to ride that wave, though -- for many the cost per GB is still too high to justify, especially if you will actually use 2 TB of disk space for your work. I'm seeing SSDs show up in greater numbers than I did two years ago -- the hard drive shortage from the floods in Vietnam had a lot of people make the jump, and once the jump was made, it's really hard to go back, especially for certain use patterns.

[Walter Roberson]

I have the new MacBook Pro 15", 8 (total) cores @ 2.6 GHz i7, NVIDIA GeForce GT 650M.

I am devoting spare compute cycles to BOINC. One of the projects I am running, Collatz Conjecture, has on OpenCL implementation that uses the GPU. Seems to work well, and appears to be doing more total work on the one GPU than the other projects are getting on 7 CPU cores combined (I leave one CPU free to handle browsing and such.)

[Walter Roberson]

The point about SSD reminds me of backup issues. The Mac Pro does not support Thunderbolt yet, and even if it did, Thunderbolt backup devices that can operate at full speed are very expensive ($3000). Most of the Thunderbolt-equipped drives on the market are only 5400 rpm, and only get a slight bit more performance than Firewire 800 can give. And Firewire 800 is rather a lot slower than SSD drives.

Whether the speed of the backup devices makes any significant difference to you will depend a lot on your use patterns and on whether you happen to have other backup infrastructure already in place.

Answer 2

[Sean de Wolski]

Is it worth the extra money to get the faster processors?

From my experience: Yes.

[Azzi Abdelmalek]

I agree with that. Faster=Much money, other considerations are subjectives

Answer 3

[Jan]

I do not use Matlab on Macs, but the experiences in this forum seems like Matlab is more stable under Windows.

I've seen Matlab programs running extremly fast on a Pentium and others hanging around stalled by false cache-sharing on quad-cores. While a SSD is fundamental for a fast start of the operating system and programs, paying a good programmer to optimize the code is a better investment than buying the hottest computer. A Terra-Giga-maschine will run 2 or 6 times faster than a standard computer, while this forum has improved the speed of some functions by a factor of 100.

Just as an example: FEX: mmx

[Paul Metcalf]

Jan, I have experience with both a HPZ800 dual processor 12 core Xeon @ 3.46GHz and an Apple single processor 6 core Xeon @ 3.33GHz. All things considered, I would recommend the Mac as specified and dual boot into Windows using Bootcamp. The OP was looking for a computer recommendation... I considered code optimisation to be beyond the scope of the question... For example, I lost 5 days of production because of a failed PS in the HP... The HP is noisier (even with liquid cooling), the Apple retains value for much longer... blah blah...

[Jan]

Sorry, Paul, you are right. My answer does miss the point that the OP wants a computer recommendation. I wanted to point out, that the "speed" does not only depend on the processor, but on the complete system including the setup, startup, design- and programming-time.

I've seen too many high-power computers wasting their time with omitted pre-allocation, and this really hurts me. A PhD student asked for a faster computer, because his optimization routines took 3 weeks. A short inspection revealed, that he has driven the internal ODE solvers with maximum accuracy all of the time. Reducing the accuracy at first and increasing it stepwise, when the optimization gets nearer to a solution, reduced the runtime to several hours.

A redundant power-supply can save days of work, a smart method for storing intermediate results on the disk also. In opposite to this, 12 cores can only run twice as fast as 6 cores.

But my preference of brain-power compared to maschine-power has an exception: 16GB RAM for 12 cores is too lean. Any disk-swapping kills efficiency, therefore it must be avoided carefully. Buy much more RAM!

Answer 4

[Paul Metcalf]

Your IT office are like any other IT office. They have no idea what they're talking about and should stick to configuring word for admin staff. You want the fastest clock rate possible + an SSD HD. You want the largest amount of on chip Level 2 and Level 3 RAM as possible. 8GB RAM is probably sufficient if you need to compromise. Unfortunately, HP sell slightly faster chips than Apple, but the Apple is a much better machine overall. What you want is this: (I'm going down the Apple Store config page) Mac Pro Single Processor Six Core Xeon @ 3.33GHz 8GB RAM 512GB Solid State HD Everything else doesn't matter. Enjoy your new machine. Paul