As an addition, it also depends on what format your source data is in.
If this is CSV, I would not hesitate using the 50 by 50 km.
For most other formats, I would suggest smaller samples.
Just a quick, very general answer: It totally depends on the size (memory footprint) of those features, and whether you can process them in one go or if FME has to hold on to them for some reason (blocking transformers).
Make sure you use 64-bit FME!
Still, it might be easier to work with smaller batches of features...