How is the compute of the model weighed against the compute that it saves?