Does adapter-based fine-tuning save memory compared to full parameter fine-tuning?

Can Adapterformer save memory usage during the training phase compared to full parameter fine-tuning? 
How much memory does full parameter fine-tuning use and how much memory does Adapterformer use?