James Tursa

@ Julio: MTIMESX does not do generic sparse matrix multiplies internally, it just calls the MATLAB built-in function mtimes. i.e., it does the same thing as if you had just typed c21 * cb at the MATLAB prompt. Can you verify that c21 * cb works at the MATLAB prompt but mtimesx(c21,cb) does not work? Be sure that you start them both from the same state (e.g., clear all other big variables including ans before attempting the calculation). That would be surprising to me since it would indicate some type of mex limitation that I am currently unaware of.

FYI, I just uploaded a new version of MTIMESX that does small matrix multiplies (first two dimensions <= 4) with inline code in 'SPEED' mode. This runs faster than the previous version.

@ D Sen: In addition, you would need to reshape audio_dft to be 32 x 1 x 4097.

Something like A(A<1e-8) = 0 should automatically remove small numbers from the sparsity pattern.

If this is too slow, have a look at spfun; it may be faster.

However, if you mean if you can remove small elements from the output matrix, you should write something of your own, partially multiplying the matrix and removing small elements. As a tip, use full column vectors of the matrix, this will make the multiplication much faster than square blocks, since sparse matrices are stored by columns. E.g.:
C = spalloc(size(A,1), size(B,2), 0);
for j=1:size(B,2)
  R = A*B(:,j);
  R(R<1e-8) = 0;
  C(:,j) = R;
end
is (probably) a fast implementation. Try to take multiple columns for speed of course.

Could it not be possible that you are really out of memory? Your result is a 460191 by 460191 matrix, which is not necessarily sparse. If the sparsity pattern is poor, the result may even be a dense matrix. Storing a full matrix this size requires 1578GiB of memory, so you have to be sure that the result is ~2% sparse. If neither of the input matrices are <1% sparse, this is likely your problem.