This is LWJGL we're talking about, the only time there's proper GPU pass-through is on linux, and even then it's kinda sketchy. So the CPU almost always has to take the brunt of the processing.
I've already tried rendering several hundred at once, and before any proper caching it's only taking about 25% of my FPS, and that's at 16 points per block. That's already less than TC's port does to me from just a few dozen blocks of track.
Most of my hit right now is that I haven't done any proper caching of vertices, so 25% hit isn't even the final number.
You don't seem to know as much about rendering as I do. ACTUAL rasterization is NEVER done CPU-side (except when you don't have a graphics card, in which case how the hell are you able to play minecraft). and 'taking 25% of my FPS' is not a good indicator of the performance hit at all, as it will vary HEAVILY per-user. Trust me, you'll want to have something better than a ton of boxes - maybe some rotated boxes lying on the spline, but not simply a bunch of axis-aligned boxes.