The Bill for Seeing Nine Pixels Across

You are designing part of a network. The block arriving at this point is 32 deep, and whatever you build must hand on a block that is also 32 deep. You have decided that each unit at the end must have a receptive field of exactly $9 \times 9$ on the block arriving here.

There are two ways to buy that field.

• Design A — a single convolutional layer with a $9 \times 9$ kernel and 32 filters.
• Design B — a stack of $3 \times 3$ convolutional layers, each with 32 filters, one feeding the next, using as many layers as it takes for the receptive field to first reach $9 \times 9$.

In both designs every layer uses stride 1 and "same" padding, so the width and height never change, and every filter carries one bias of its own.

How many more learnable parameters does Design A have than Design B?

Count weights and biases together, across every layer of each design. Give the exact integer.