Memory efficient multiplication of large matrices on GPU

Dear community,

I would like to use the GPU enabled functions of MATLAB to perform quick matrix multiplications. I already discussed the fundamental idea here: https://uk.mathworks.com/matlabcentral/answers/363627-gpu-enabled-functions-react-slow-on-first-call

Since gDexMat is very big, the GPU quickly runs out of memory. However, gDexMat itself can be calculated on the fly from vectors as follows:

>> x0=400e3;
>> Xs=200e6;
>> n=500;

>> a=gpuArray((x0:x0:Xs/2));
>> b=gpuArray((1/Xs:1/Xs:n/x0)');
>> gDexMat=cos(a*b);
>> whos a b gDexMat
Name              Size                  Bytes  Class     Attributes

a                 1x250                  4     gpuArray
b              250000x1                  4     gpuArray              
gDexMat       50000x250                  4     gpuArray

The desired operation is:

 >> whos Sig
  Name        Size                    Bytes  Class     Attributes

Sig       100x250000                  4    gpuArray

>> Result=Sig*gDexMat;
>> whos Result
Name          Size              Bytes  Class     Attributes

Result      100x250               4    gpuArray

which could of course also be written as:

 >>  Result=Sig*cos(a*b);

In this case however, it appears as if the result of cos(a*b) is calculated first and than multiplied with Sig, causing the GPU to run out of memory again. Even switching to single precision did not help much.

Is there an efficent way to perform this operation on the GPU? Perhaps, the "funarray" functions could solve this issue? If so, could somebody provide me with an example on how to do so?

Thanks for your help,

Ludwig