I think this multiplication process was dealt with at the last lecture. By breaking up the bigger matrix multiplication into smaller sub multiplications. From the last optional assignment, we can see that his approach makes it easier to create parallel work for matrix multiplication. This operation seems efficient because this minimizes the movement of data and eventually the energy loss.

Does this strategy for building larger matrix-matrix multiplies work for matrices that are so big, there is no hope of storing even a single row of the matrix in the cache? I wonder if there are other strategies for partial evaluation of matrices (e.g. where you only need to load a constant number of elements into the cache at a time).