Matlab R2014b CPU and GPU Matrix Multiply Time Comparison

Matlab has incorporated GPU processing on the parallel computing toolbox and you can create GPU array objects using the gpuArray(…) function in MATLAB. I created a brief script to compare  matrix multiply of a 2048 x 2048 matrix against a vector. Ordinarily, the CPU operations see reasonable speedup (~2x) from moving from double to single precision values. However, moving to the GPU implementation results in a speedup of  6.8x for Double and  5.6x for Single! This means that if you can take a matrix-vector multiply that is double precision and convert it to single precision GPU version, you may see  a gain of nearly 14x.

The following we generated in Matlab R2014b on an i7-4770 3.5 Ghz CPU  (8 CPUs) with 16GB Ram and a Geforce GTX 750.

The next step is to evaluate speed of the gpuArray on a basic L1 Optimization set—l1 magic.

The code used to generate this data is as follows:

 

allTypes = {'Double', 'gpuArrayDouble', 'Single', 'gpuArraySingle'};
allTimes = nan(length(allTypes),1);

n = 2048;       % size of operation for Ax
num_mc = 2^10;  % number monte carlo runs to compute time average of runs

randn('seed', 1982);
Am = randn(n,n);
xm = randn(n,1);      

for ind_type = 1:length(allTypes)
   
    myType = allTypes{ind_type};
     
    switch lower(myType)
        case 'double'
            A = double(Am);
            x = double(xm);
        case 'single'
            A = single(Am);
            x = single(xm);
        case 'gpuarraydouble'
            A = gpuArray(Am);
            x = gpuArray(xm);
        case 'gpuarraysingle'
            A = gpuArray(single(Am));
            x = gpuArray(single(xm));
        otherwise
            error('Unknown type');            
    end    
    
    tic
    for ind_mc = 1:num_mc
        y = A*x;
    end
    allTimes(ind_type) =  toc/num_mc;
    
end

%% Display the results
figure(34);
clf;
bar(allTimes*1000);
set(gca, 'xticklabel', allTypes, 'color', [1 1 1]*.97);
title(['Timing of CPU and GPU in M' version('-release')]);
xlabel('Type');
ylabel('Time (ms)');
grid on

%%
figure(35);
clf;
speedupLabels = {('Double to Single') , ...
                 ('Double to gpuDouble'), ...
                 ('Single to gpuSingle'), ...
                 'Double to gpuSingle'};
bar([allTimes(1)/allTimes(3), allTimes(1)/allTimes(2),  ...
    allTimes(3)/allTimes(4), allTimes(1)/allTimes(4)]);
set(gca, 'xticklabelrotation', 15, 'xticklabel', speedupLabels, 'color', [1 1 1]*.97);
title(['Speedup of CPU and GPU in M' version('-release')]);
xlabel('Type');
ylabel('Speedup');
grid on