Time series prediction using neural networks (narnet / narxnet); is it at all possible to train a network faster on a GPU than the CPU?

I have gpuBenched ( https://www.mathworks.com/matlabcentral/fileexchange/34080-gpubench ) my Geforce GTX 570 GPU, in this calculation the GPU proved to be orders of magnitude faster than the CPU.

Left: Results for data-type 'double' (In GFLOPS)

Right: Results for data-type 'single' (In GFLOPS)

                           MTimes   Backslash  FFT     MTimes  Backslash  FFT
  GeForce GTX 570     175.11   103.31     60.14   954.68  287.08   187.61
  Host PC             41.02    35.08      7.13    82.34   66.76     11.61

But...When I use the GPU to train a neural network it takes much longer time compared to training with the CPU. For the code below it takes 20 times longer using the GPU. I have tried different sizes of dataset, different feedback lengths and different amount of neurons but the conclusion remains.

Any suggestions of how the GPU can be used more efficiently for these types of time series prediction calculations?

Have you preformed calculation involving time series prediction were the GPU was faster than the CPU?

Also; Do you know why Mathworks hasn't developed the Jacobian training methods (eg. trainlm, trainbr (these are the best ones)) for GPU training/computing?

The code (paste and run, this is an executable syntax):

clear all
close all
g = gpuDevice(1);
reset(g);

FD=1:50;
T = solar_dataset;
T=[T T T T T T T T T T];
neu=10;

%%GPU CODE

tic
rng('default');

   for Ntrials=1:3

net = narnet(FD,neu);

    net.trainFcn ='trainscg';

for i=1:net.numLayers
  if strcmp(net.layers{i}.transferFcn,'tansig')
    net.layers{i}.transferFcn = 'elliotsig';
  end
end

[Xo,Xoi,Aoi,To] = preparets(net,{},{},T);   
net = configure(net,Xo,To);
Xo = nndata2gpu(Xo,'single'); 
To = nndata2gpu(To,'single'); 
Xoi = nndata2gpu(Xoi,'single'); 
Aoi = nndata2gpu(Aoi,'single');

[ net] = train(net,Xo,To,Xoi,Aoi);
Yo = net(Xo,Xoi,Aoi);

Yo=gpu2nndata(Yo);
To=gpu2nndata(To);
Eo = gsubtract(To,Yo);

clear net
reset(g);

Ntrials

   end

toc

%%Removing some variables to make everything equal

clearvars -except FD T neu

   %%CPU CODE

tic
rng('default');

   for Ntrials=1:3

net = narnet(FD,neu);

    net.trainFcn ='trainscg';

for i=1:net.numLayers
  if strcmp(net.layers{i}.transferFcn,'tansig')
    net.layers{i}.transferFcn = 'elliotsig';
  end
end

[Xo,Xoi,Aoi,To] = preparets(net,{},{},T);   
net = configure(net,Xo,To);
% Xo = nndata2gpu(Xo,'single'); 
% To = nndata2gpu(To,'single'); 
% Xoi = nndata2gpu(Xoi,'single'); 
% Aoi = nndata2gpu(Aoi,'single');

[ net] = train(net,Xo,To,Xoi,Aoi);
Yo = net(Xo,Xoi,Aoi);

% Yo=gpu2nndata(Yo);
% To=gpu2nndata(To);
Eo = gsubtract(To,Yo);

clear net
% reset(g);

Ntrials

   end

toc

[Staffan]

Small update, I have also tried gpuArray() instead of nndata2gpu() but this is only slightly faster and does not give the sought after level of improvement.

For some reason gpuArray() also seems to be way more inaccurate (= larger NMSEo) than nndata2gpu().

[Staffan]

Anyone? Any help on this subject is highly appreciated

[Ricardo de Azevedo]

Hi Staffan, I am looking for the answer to the same question, and found this thread that has been helpful:

https://www.mathworks.com/matlabcentral/answers/225961-gpu-training-of-neural-network-with-parallel-computing-toolbox-unreasonably-slow-what-am-i-missing

Did you find a solution to the problem?

[Staffan]

Hello Ricardo,

I actually generated a support case for this, the answers I got from Mathworks was nothing to cheer for. After said responses I started looking at Tensorflow, I find this tool much more appropriate for what I'd like to accomplish...I've worked with univariate time series and did not get the performance from Narnet that I thought I was going to have.

BR Staffan