Why GPU-parallel simulation changes RL training economics

Frameworks like NVIDIA Isaac Lab (successor to Isaac Gym) run thousands of environment instances in parallel directly on the GPU, keeping observations, the physics step, and the policy all on-device. Compared to a traditional CPU-based simulator (e.g., one or a handful of Gazebo instances) feeding a GPU that only runs the neural network, what is the primary reason this GPU-native, massively-parallel approach dramatically speeds up reinforcement-learning training?