Making learning on the edge work requires a significant reduction in power, which means rethinking all parts of the flow.
