It is possible to enable/disable the CNN engine by 'cnn_enable()' and 'cnn_disable()' functions. Once the clock has stopped, this engine does not consume energy. It is not necessary to re-initialize the model between cnn_disable() & cnn_enable(). However, these functions cannot interrupt the model inference. To pause the model during inference, one must split the model into parts (layer-by-layer if necessary), load each model part into kernel memory, and manage the inference properly.
The different operating modes and clock settings are described in more details in the following article: https://www.analog.com/en/resources/app-notes/developing-power-optimized-apps-on-the-max78000.html