The choose of number of environments to maximize the performance

Hello,

I tried the Preview 3 with Quadcopter task.
If num_envs is 8192, the per hit on self.gym.simulate(self.sim) took around 8 ms.
If num_envs is 2048, the per hit on self.gym.simulate(self.sim) took around 4 ms.
If num_envs is 512, the per hit on self.gym.simulate(self.sim) took around 3 ms.

I thought 8192 envs will take the same amount of time as 2048 envs to complete the simulation. Does it seem the simulation is not a fully parallel process?

My question is that how does the parallel simulation process work? Is that larger num_envs mean higher speed overall? I see the portion (% time) of pre_physics_step and post_physics_step decrease as num_envs increase.

Thank you!

Best,