The core idea of device-resident simulation is to keep both the simulation state and the computation graph resident in GPU memory, so that the entire simulation-to-inference closed loop is completed entirely within the device, synchronized only through kernel launches, thereby eliminating copy overhead.