The paper cited is about hardware, where there is no accuracy tradeoff because you control the numerical precision completely and use fixed point. In a software implementation, neither is true. There is no chance that you will get the exact same values out of this method that you do out of other FP matmuls.