Why for loop in CPU faster than in GPU

I am running the following command:

p = zeros(7000000,1); g_p = gpuArray(p);

1. CPU version

for i=1:size(p,1)
p(i) = 10; % just simple to see
end

2. GPU version

for i=1:size(g_p,1)
g_p(i) = 10; % just simple to see
end

It is amazing that the execution time in CPU really more faster than it in GPU. Why?

[Adam]

Data has to be copied to the GPU. This has an overhead so you use the GPU for massively parallel intensive computations rather than things that are so fast on the CPU that the overhead of copying to the GPU is greater than the saving made.

[Jan]

@Khieu: Compare this:

tic; p(:) = 10; toc
tic; g_p(:) = 10; toc

What do you get?

[Khieu]

Thanks Adam

[Adam]

It is very difficult often to determine where this crossover is. I have occasionally dabbled with a GPU implementation of some computation in my code, but it is always hit and miss whether it is faster or slower on a particular run since data sizes vary so I often just end up commenting it out with a vague idea I might come back to it one day and consider some fancy code to chose the CPU or GPU version of my code based on input size and any other appropriate metrics, but I never have done so far.

Answer 1

[Joss Knight]

This isn't amazing at all. You are launching 7 million kernels, each to copy one value into GPU memory. The kernels cannot overlap because they are operating on the same data, which means each must wait for the last to finish. This is not at all an efficient way to do parallel programming!