The Same Filters, a Bigger Photo

A convolutional layer is defined once, with these settings:

It is then used twice. First it runs on a $32 \times 32 \times 8$ block. Later the very same layer — the same filters, holding the same trained numbers — runs on a $64 \times 64 \times 8$ block.

For each of the two runs, count two things:

• the learnable weights the layer holds (ignore the bias terms),
• the multiplications one forward pass performs — one multiplication for every pixel-times-weight product, ignoring the additions.

Which pair of figures is right?