The exact simulation scheme by Broadie & Kaya (2006) is a ground-breaking research piece, but the algorithm is slow and complicated to implement. IMHO, that's why it is not so popular among practitioners.
The algorithm is "exact" in the sense that you can 'jump' any arbitrary time step as opposed to the time-discretization (e.g., Euler/Milstein or Andersen (2008)'s QE) scheme where you have to jump a small time step (therefore, many jumps) in order to control bias. In the exact scheme, however, you have to sample the integrated variance from given the terminal variance. This step is quite slow because the distribution of the integrated variance is given by its Fourier transform (i.e., characteristic function), so you need to numerically invert to obtain the CDF. Caching the CDF is not a viable option here because the CDF is unique given the terminal variance. So the quick0-and-dirty time-discretization scheme is still a practical solution in terms of performance and implementation effort. Lord et al (2010) and Van Haastrecht & Pelsser (2010) have some comparison results including the exact scheme, so please take a look.
But there is another important algorithm improving the drawback of the exact scheme. Glasserman & Kim (2011) express the integrated variance as the infinite sums of gamma random variables, removing the bottleneck in the exact scheme. Of course, you can't do the infinite sums in numerical implementation. But, as long as you use a reasonable number (e.g., <10) of terms, the truncation error is quite small.