In Simple convolutional Network at 3:35minutes, after strided convolution we get 17 x 17x 20(16.5 is rounded up). I remember in slide 17, we were told to round down if it is in decimal point(floor(z)).

Does it really matter?

Yes, it matters. The formula for the output dimension is actually quite explicit: you round down (“floor”) and then add 1.

n_{out} = \displaystyle \lfloor \frac {n_{in} + 2p - f}{s} \rfloor + 1

Where f is filter size, p is padding and s is stride.