Modern Intel architectures supporting fma instruction sets can perform the first loop calculating the matrix-matrix product between panels a and b in one go using _mm256_fmadd_pd. We should implement these and see how they affect performance.