What do they mean by 'To calculate sample size, I use simulation in all cases.'? I was looking for a book on sample size calculations and power analysis and I met this phrase 'I use simulation in all cases.'.
 What does it mean, can i forget about using traditional methods for calculating sample size and start learning  simulation ?
 A: I don't know how you'd infer any involvement of Bayesian calculations from that link.
Presumably by "they" you mean myself and Stephan (since you linked to my answer but he's the one that said what you're asking about). I said that I tend to use simulation for any but the simplest cases. Stephan said he'd use it in all cases.  
Note that the situation being described at that linked question was reasonably complex so in that case it is one we'd both be using simulation on.
In complicated situations, some people (and quite a few books) fall back on simplifications or other approximations. I lean toward simulation since it's easy to apply and doesn't rely on approximating the model (though the answers from the simulation are themselves approximate, you can make them more accurate by simulating more). 
Such simulations work by choosing some alternative and simulating data from the model at that alternative to obtain the rejection rate (i.e. the power) at some sample size; often you can then work out the required sample size (but you can update it by trying that calculated sample size in another simulation). Alternatively if you want to investigate the power under a range of effect sizes (which I would prefer to a single-point calculation, since the effect size won't be the one supposed in the calculation of sample size), you can simulate power curves or sets of power curves.
In addition a slight modification of the simulation will let you see what happens to power and significance levels as assumptions fail to hold in whatever manner and degree you wish to investigate (so you can see how sensitive things are to heteroskedasticity, dependence, different distributional assumptions and other forms of model misspecification).
This doesn't really mean you can "forget" traditional methods, because if you're going to simulate the right things, it's important to understand what you're doing;  the traditional calculations (not knowing the formulas, understanding how they were obtained) would be an essential part of the grounding (as is understanding simulation fairly well).
Once you have some practice at it, simulations a very useful tool, but it needs a foundation of understanding to build on.
It will let you solve problems that are hard to solve other ways. But how would you know you didn't make a mistake? If you know how to approximate the right answer analytically, you have access to an important check on your simulation. 
. 
