Is it? I could not see from the paper whether they assume non-commutative multiplication in their small matrix optimization.

//Edit: they do a 4x4 block matrix, but the gains are less than 5% over the existing Strassen algorithm.