#efficiency

A convolution can be written as $\mathbf{y} = \mathbf{W}\mathbf{x}$, where $\mathbf{x}$ is the flattened input and each row of $\mathbf{W}$ holds the kernel weights for one output position and zeros elsewhere. For a $32 \times 32$ single-channel input and a $3 \times 3$ kernel (stride 1, no padding), $\mathbf{W}$ has $900 \times 1024$ entries. What percentage of them are nonzero?