VBO size doubts in weird testing results

I’m working on visualization of massive point clouds, somewhere in the ballpark of 330 million points per scene. I use VBOs to upload this data to VRAM and I couldn’t find a lot of info on what the recommended size of each VBO is.

Only thing I could find was an old nvidia document that mentioned 1-4 MB per VBO was recommended. Is this still valid?

I also tested loading times with different VBO sizes, with results that I didn’t expect:


This is the load time depending on the VBO size for 700K points, this is not the load time of a single VBO its the total time needed to load as many VBOs needed to fit the 700K points.

I thought that the graph was going to be smoother without that much difference for more or less the same size.

Is this normal? Why does this happen?