yeah of course it depends alot on the data and how complex the model is. What would be nice to see is an example of someone that managed to stretch this way further performing bayesian inference on an highly nonlinear very complex model with the NUTS-sampler in pymc. Most people i´ve seen that worked on very complex models with potentially large datasets tend to choose SVI from what i´ve seen. Btw, i pm:ed you about your article.