SGD is fast but not necessarily more accurate. If you've got lots and lots of data, then a simple yet fast approach is likely to be a good choice to start with.

Thanks.

I also found a blog post that figured the diagram, with some background infos:

http://peekaboo-vision.blogspot.de/2013/01/machine-learning-...

And a prior HN discussion: https://news.ycombinator.com/item?id=5831512

There are many others too, but with few points and comments