I'm doing some tests again: 300 units in 10 groups are moving smoothly, but for 600 units in 20 groups, with pssm shadows, I got now about 14 fps when all is moving (when standing, it is over 40). Vertex animation is used, it's not as bad as before, if heavy lod applied: 400, 250, 125, and 75 tris.
I found it faster to use pssm than decal shadows, over a certain quantity.

maybe I should use physx instead of c_move to get better performance?