Interesting results... my gut instinct would be that the bottleneck is RAM, then... it's usually easier to congest your RAM than your CPU these days. Most of the overhead is the sheer number of instances, not the heavy computations they do (some basic trigonometry and movement) so that checks out as a RAM issue in theory as well. CPU-->GPU transfer might be an issue since I don't really buffer anything particularly optimizedly, but the drawing code is pretty basic so I'm more inclined to think it's RAM transfer speeds.