As I asked before in https://www.mathworks.com/matlabcentral/answers/1911930-how-to-calculate-number-of-floating-point-operations-flops-of-a-cnn-model-in-matlab#comment_2614255 we need to know whether you are asking about speed of training or speed of application after training ?
Does this CNN live entirely on the CPU, or does it also live partly on the GPU ?
GPU measurement is tricky because different models of GPU can implement some operations quite differently. In particular, the NVIDIA models can have very different double precision performance than other models.
