I understand that by Cholesky Decomposition, multivariate normal distribution $X=[X_1, ..., X_n]$~$ N(0,\Sigma)$ can be simulated as $RZ$, where $Z=[Z_1, ..., Z_n]$~$N(0,I_{n\times n})$ and $R$ is a lower triangular matrix such that $RR^T=\Sigma$.
My question is:
Assume that for each variable $X_i$, there are $m$ samples. Notice that $\Sigma_{ij} =cov(X_i,X_j)=\frac{1}{m}\sum_{k=1}^{m}(x_{i,k}-\overline{x}_i)(x_{j,k}-\overline{x}_j)=\sum_{k=1}^{m}y_{i,k}\times y_{j,k}$, where $y_{i,k} = \frac{x_{i,k}-\overline{x}_i}{\sqrt{m}}$. Therefore, we can define a matrix $A_{n\times m}$ with $a_{ij}=y_{ij}$. It's easy to see that
$AA^T=\Sigma$
So $X$ can be simulated as $AZ'$, where $Z'=[Z_1,...,Z_m]$~$N(0,I_{m\times m})$. Why don't people simply use this method to simulate multivariate normal distribution (by avoiding the computation to find $R$)? What's the disadvantage of this method compared to Cholesky Decomposition?