why use gpu training the performance is worse than cpu

Hi,

I tried the following options for training a fitnet:

trainedNet=train(net,X_',Y'); %59 Iterations,Performance 115, 100s

trainedNet=train(net,X_',Y','useParallel','yes');%59 Iterations ,Performance:115,158s

trainedNet=train(net,gpuArray(X_'),gpuArray(Y')); %628 Iterations,Performance:649, 14s

trainedNet=train(net,X_',Y','useGPU','yes');%628 Iterations,Performance:679, 12.6s

trainedNet=train(net,X_',Y','useGPU','only');% 628 Iterations,Performance:679, 12.6s

I fixed the random seed.

I want to know why using GPU it took much more iterations but achieve worse performance?

Thanks

[Yunyu Hu]

I found the major reason, because I used the trainfcn "trainlm", which is not supported by GPU and when use GPU matlab will change it automatically to "trainscg".

But then it comes the next question: when I use " trainscg" for bothe CPU and GPU, the Performances are different:

trainedNet=train(net,X_',Y','useGPU','only‘) --> 5000Itr,Performance 192, 232s

trainedNet=train(net,X_',Y‘); --> 5000Itr, Performance 154, 186s

What is the reason for this?

Answer 1

[Sai Bhargav Avula]

Hi,

First I would like to mention that GPU's are faster for large matrices.

I would also like to suggest two answers that explains GPU performance

https://www.mathworks.com/matlabcentral/answers/385598-why-gpuarray-is-slower-than-cpu

https://www.mathworks.com/matlabcentral/answers/26552-gpu-time-slower-than-cpu-time-what-went-wrong-with-my-gpu-implementation

Hope this helps!