Question about Convolution on GPU

[JAEWOOKe]

Dear all,

I have been running ResNet and GoogLeNet on GPU and the first convolution layer of each model costs a lot of time.

This also leads to the result that GPU is much slower than CPU.

For ResNet50, I got the below result.

                     CPU               GPU
Conv      |    29.462    |    186.912  |
Other     |     5.134     |      2.063    |

Can anyone give me a piece of advice on what's happening?

Thanks.