Total CPU/GPU time for COCA with varying numbers of force evaluations, n_f, compared to COLA. Both COCA and COLA use 20 time steps. In the left column, we report the total time, and in the right column, we report the relative contributions of the various operations. Solid bars correspond to the main computations, and hashed bars indicate inputs and outputs. We note that the current implementation of COCA in SIMBELMYNë decouples emulation and other computations, which is not a requirement–emulation could occur on the fly during N-body evolution. In this way, the cost of input+output operations would be effectively reduced to zero.