How to know in advance if creating matrix will run out of memory?

I need to create a matrix of size N-by-M as part of a script, where N and M are inputs to the script.

So I will know N and M, and then depending on their size choose do one of the following three:

i) I want to create a gpu matrix when N and M are small enough that the relevant matrix will fit on gpu without creating an out of memory error.

ii) When N and M are too large for this I want to create the matrix on the cpu instead.

iii) If N and M are so large that I would run out of memory on the cpu then I want to create the matrix as a sparse matrix on the cpu.

Is there a way, once I know N and M, to check if the matrix will be possible to create?

[I could use 'try-catch-end', with try just erroring when creating a matrix too large for memory, but I would prefer something cleaner as I suspect that runtime to do the try-catch on a matrix too large for memory is wasteful]

Answer 1

[Raymond Norris]

Look at the memory command, but it's only for Windows operating system. The gpuDevice command will return AvailableMemory of a GPU. This can get you started. Keep in mind memory needed to operate on it (e.g., adding matrices, fft, etc.). It might suprise you how much memory gets used on a GPU with several matrices.

[Robert Kirkby]

Thanks.

I'm on linux so memory won't work. But checking gpu memory is much more important to me so that will help.

Is there an easy way to tell how much memory X will be used in advance?, e.g., that a N-by-M double floating point matrix will use X gigabytes?

[Raymond Norris]

Try

sz = N * M * size-of-unit / 1024 ^ units;

size-of-unit depends on int8 . . . int64 (i.e., 1 . . . 8) and units depends on MB or GB (i.e., 2 or 3). For example, size of N,M double matrix in GBs would be

sz = N * M * 8 / 1024 ^ 3;