Minimizing Inner Product Data Dependencies in Conjugate Gradient Iteration