MATLAB and Simulink resources for Arduino, LEGO, and Raspberry Pi

Learn moreOpportunities for recent engineering grads.

Apply TodayTo resolve issues starting MATLAB on Mac OS X 10.10 (Yosemite) visit: http://www.mathworks.com/matlabcentral/answers/159016

Asked by C Zeng on 17 Mar 2013

Hi, All,

I have to use a dataset as large as ones(3^N1,3^N2,10^3), where N1 and N2 can be 7~14. However, matlab will report "Out of memory" even for ones(3^14,1000). Is there a way to let matlab store such a large array.

I can split my data to several parts and store the sub-data. However that is very inconvenient.

I wonder if MATLAB can extend the limit, and handle large scale data. Thanks.

*No products are associated with this question.*

Answer by Walter Roberson on 17 Mar 2013

Edited by Walter Roberson on 17 Mar 2013

Accepted answer

If you are able to use 8 bits per value, you could use

ones(3^N1, 3^N2, 10^3, 'uint8') %or 'int8' as appropriate.

That would occupy 4.45 gigabytes, which would be too much for any 32 bit version of MATLAB (32 bits is 4 gigabytes of address space.) It would, however, be feasible with a 64 bit MATLAB with as little as 8 gigabytes of RAM.

If you need to use double precision, then that occupies 8 bytes per location, and so requires 35.64 gigabytes for the array. You would need a 64 bit MATLAB for sure, and you would need probably 40 or so gigabytes of RAM. If that much RAM is not feasible for you, then using a double precision array of that size is not feasible for you.

This is not a matter of how MATLAB uses its memory: this is just a matter of having enough memory to store the array at all.

Note: your options are different if the array is reasonably sparse.

Show 4 older comments

C Zeng on 19 Mar 2013

Yes, Cedric,

Thanks for showing the details. I have a 4GB RAM, now I see why Matlab reports out of memory. Let me work on server to see if it allows me to use 10GB RAM.

Cedric Wannaz on 19 Mar 2013

But be aware that `uint8` can store only 8 bits unsigned integers, which means integers in the range 0-255. It might not be adapted to the nature of your data, which might lead to e.g. truncation..

>> a = uint8(200) a = 200 >> 2*a % Will it be 400? ans = 255

I can say by experience that often, when people would need much more RAM than what is present on an average computer to perform their computation, the problem is not that computers don't have enough RAM, but that the approach is flawed.

Answer by Cedric Wannaz on 17 Mar 2013

Edited by Cedric Wannaz on 17 Mar 2013

If your data is sparse, you can build 1000 sparse matrices or use some FEX function that will allow you to create ND sparse matrices.

The `ones(3^14, 1000)` that you tried to evaluate would take more than 38GB RAM to be stored as `double`. It certainly indicates that your approach is not appropriate. One thing that you could do though, is to store your data using a "smaller" type/class, i.e. any decent enough type (for the nature of your data) that is stored on 1, 2, or 4 bytes. If your data were made of unsigned integers in the range 0-255, you could use `uint8` that would require 1 byte per element instead of 8.

## 0 Comments