You're right that it's tricky to make a direct comparison between different shaped optimization domains, however in the comparison animation each algorithm was given an optimization domain of equal area. For comparisons in higher dimensions it is up to the user to specify optimization domains with equal content.
Simple's runtime performance is not affected by the size of the optimization domain, however any global optimization algorithm will necessarily need to make more samples to chraracterize a higher dimensional space because these spaces inherently have more content to explore. Essentially the runtime performance of the algorithm and its sample-efficiency are two orthogonal concepts. Simple promises to be at least as sample-efficient as Bayesian Optimization while offering significantly lower computational overhead. Its main selling point isn't improved sample efficiency so much as better performance.
Does that answer your questions?
EDIT: Regarding the appropriate problem dimensionality, that will tend to depend quite a bit on the specific optimization problem. Simple is a global optimizer, so for extremely dimensional problems like training a deep neural net it would not be appropriate because the optimization domain would be too massive. The algorithm needs more testing, but as a rule of thumb I would say that anything under 500 dimensions is feasible.
If you take a simplex of length size d as starting simplex so that it contains the unit cube, does it still beat Bayesian Optimization ? Or maybe map the simplex to the cube but it will distort things a lot.
What is a typical value for the dimension of the space where your algorithm should perform ?