Not if you have duplicates of rows on the first matrix, which can be done very efficiently if you build specialized hardware. Then its all just forward in parallel.