Each square is a whole tile (think 128×128 numbers). One step = one block-product: fetch a tile of A and a tile of B, multiply, accumulate into the output patch. A patch needs all 4 of its inner-index terms before it is finished — then it ships to the library and its desk space is reused. The small corner numbers on A and B count how many times each tile has been fetched - watch the reuse story return at block level.